(Another) minor refactoring of substitutions
[ghc.git] / compiler / simplCore / Simplify.hs
1 {-
2 (c) The AQUA Project, Glasgow University, 1993-1998
3
4 \section[Simplify]{The main module of the simplifier}
5 -}
6
7 {-# LANGUAGE CPP #-}
8
9 module Simplify ( simplTopBinds, simplExpr, simplRules ) where
10
11 #include "HsVersions.h"
12
13 import DynFlags
14 import SimplMonad
15 import Type hiding ( substTy, substTyVar, extendTvSubst, extendCvSubst )
16 import SimplEnv
17 import SimplUtils
18 import FamInstEnv ( FamInstEnv )
19 import Literal ( litIsLifted ) --, mkMachInt ) -- temporalily commented out. See #8326
20 import Id
21 import MkId ( seqId, voidPrimId )
22 import MkCore ( mkImpossibleExpr, castBottomExpr )
23 import IdInfo
24 import Name ( Name, mkSystemVarName, isExternalName )
25 import Coercion hiding ( substCo, substCoVar )
26 import OptCoercion ( optCoercion )
27 import FamInstEnv ( topNormaliseType_maybe )
28 import DataCon ( DataCon, dataConWorkId, dataConRepStrictness
29 , isMarkedStrict, dataConRepArgTys ) --, dataConTyCon, dataConTag, fIRST_TAG )
30 --import TyCon ( isEnumerationTyCon ) -- temporalily commented out. See #8326
31 import CoreMonad ( Tick(..), SimplifierMode(..) )
32 import CoreSyn
33 import Demand ( StrictSig(..), dmdTypeDepth, isStrictDmd )
34 import PprCore ( pprCoreExpr )
35 import CoreUnfold
36 import CoreUtils
37 import CoreArity
38 --import PrimOp ( tagToEnumKey ) -- temporalily commented out. See #8326
39 import Rules ( mkRuleInfo, lookupRule, getRules )
40 import TysPrim ( voidPrimTy ) --, intPrimTy ) -- temporalily commented out. See #8326
41 import BasicTypes ( TopLevelFlag(..), isTopLevel, RecFlag(..) )
42 import MonadUtils ( foldlM, mapAccumLM, liftIO )
43 import Maybes ( orElse )
44 --import Unique ( hasKey ) -- temporalily commented out. See #8326
45 import Control.Monad
46 import Outputable
47 import FastString
48 import Pair
49 import Util
50 import ErrUtils
51
52 {-
53 The guts of the simplifier is in this module, but the driver loop for
54 the simplifier is in SimplCore.hs.
55
56
57 -----------------------------------------
58 *** IMPORTANT NOTE ***
59 -----------------------------------------
60 The simplifier used to guarantee that the output had no shadowing, but
61 it does not do so any more. (Actually, it never did!) The reason is
62 documented with simplifyArgs.
63
64
65 -----------------------------------------
66 *** IMPORTANT NOTE ***
67 -----------------------------------------
68 Many parts of the simplifier return a bunch of "floats" as well as an
69 expression. This is wrapped as a datatype SimplUtils.FloatsWith.
70
71 All "floats" are let-binds, not case-binds, but some non-rec lets may
72 be unlifted (with RHS ok-for-speculation).
73
74
75
76 -----------------------------------------
77 ORGANISATION OF FUNCTIONS
78 -----------------------------------------
79 simplTopBinds
80 - simplify all top-level binders
81 - for NonRec, call simplRecOrTopPair
82 - for Rec, call simplRecBind
83
84
85 ------------------------------
86 simplExpr (applied lambda) ==> simplNonRecBind
87 simplExpr (Let (NonRec ...) ..) ==> simplNonRecBind
88 simplExpr (Let (Rec ...) ..) ==> simplify binders; simplRecBind
89
90 ------------------------------
91 simplRecBind [binders already simplfied]
92 - use simplRecOrTopPair on each pair in turn
93
94 simplRecOrTopPair [binder already simplified]
95 Used for: recursive bindings (top level and nested)
96 top-level non-recursive bindings
97 Returns:
98 - check for PreInlineUnconditionally
99 - simplLazyBind
100
101 simplNonRecBind
102 Used for: non-top-level non-recursive bindings
103 beta reductions (which amount to the same thing)
104 Because it can deal with strict arts, it takes a
105 "thing-inside" and returns an expression
106
107 - check for PreInlineUnconditionally
108 - simplify binder, including its IdInfo
109 - if strict binding
110 simplStrictArg
111 mkAtomicArgs
112 completeNonRecX
113 else
114 simplLazyBind
115 addFloats
116
117 simplNonRecX: [given a *simplified* RHS, but an *unsimplified* binder]
118 Used for: binding case-binder and constr args in a known-constructor case
119 - check for PreInLineUnconditionally
120 - simplify binder
121 - completeNonRecX
122
123 ------------------------------
124 simplLazyBind: [binder already simplified, RHS not]
125 Used for: recursive bindings (top level and nested)
126 top-level non-recursive bindings
127 non-top-level, but *lazy* non-recursive bindings
128 [must not be strict or unboxed]
129 Returns floats + an augmented environment, not an expression
130 - substituteIdInfo and add result to in-scope
131 [so that rules are available in rec rhs]
132 - simplify rhs
133 - mkAtomicArgs
134 - float if exposes constructor or PAP
135 - completeBind
136
137
138 completeNonRecX: [binder and rhs both simplified]
139 - if the the thing needs case binding (unlifted and not ok-for-spec)
140 build a Case
141 else
142 completeBind
143 addFloats
144
145 completeBind: [given a simplified RHS]
146 [used for both rec and non-rec bindings, top level and not]
147 - try PostInlineUnconditionally
148 - add unfolding [this is the only place we add an unfolding]
149 - add arity
150
151
152
153 Right hand sides and arguments
154 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
155 In many ways we want to treat
156 (a) the right hand side of a let(rec), and
157 (b) a function argument
158 in the same way. But not always! In particular, we would
159 like to leave these arguments exactly as they are, so they
160 will match a RULE more easily.
161
162 f (g x, h x)
163 g (+ x)
164
165 It's harder to make the rule match if we ANF-ise the constructor,
166 or eta-expand the PAP:
167
168 f (let { a = g x; b = h x } in (a,b))
169 g (\y. + x y)
170
171 On the other hand if we see the let-defns
172
173 p = (g x, h x)
174 q = + x
175
176 then we *do* want to ANF-ise and eta-expand, so that p and q
177 can be safely inlined.
178
179 Even floating lets out is a bit dubious. For let RHS's we float lets
180 out if that exposes a value, so that the value can be inlined more vigorously.
181 For example
182
183 r = let x = e in (x,x)
184
185 Here, if we float the let out we'll expose a nice constructor. We did experiments
186 that showed this to be a generally good thing. But it was a bad thing to float
187 lets out unconditionally, because that meant they got allocated more often.
188
189 For function arguments, there's less reason to expose a constructor (it won't
190 get inlined). Just possibly it might make a rule match, but I'm pretty skeptical.
191 So for the moment we don't float lets out of function arguments either.
192
193
194 Eta expansion
195 ~~~~~~~~~~~~~~
196 For eta expansion, we want to catch things like
197
198 case e of (a,b) -> \x -> case a of (p,q) -> \y -> r
199
200 If the \x was on the RHS of a let, we'd eta expand to bring the two
201 lambdas together. And in general that's a good thing to do. Perhaps
202 we should eta expand wherever we find a (value) lambda? Then the eta
203 expansion at a let RHS can concentrate solely on the PAP case.
204
205
206 ************************************************************************
207 * *
208 \subsection{Bindings}
209 * *
210 ************************************************************************
211 -}
212
213 simplTopBinds :: SimplEnv -> [InBind] -> SimplM SimplEnv
214
215 simplTopBinds env0 binds0
216 = do { -- Put all the top-level binders into scope at the start
217 -- so that if a transformation rule has unexpectedly brought
218 -- anything into scope, then we don't get a complaint about that.
219 -- It's rather as if the top-level binders were imported.
220 -- See note [Glomming] in OccurAnal.
221 ; env1 <- simplRecBndrs env0 (bindersOfBinds binds0)
222 ; env2 <- simpl_binds env1 binds0
223 ; freeTick SimplifierDone
224 ; return env2 }
225 where
226 -- We need to track the zapped top-level binders, because
227 -- they should have their fragile IdInfo zapped (notably occurrence info)
228 -- That's why we run down binds and bndrs' simultaneously.
229 --
230 simpl_binds :: SimplEnv -> [InBind] -> SimplM SimplEnv
231 simpl_binds env [] = return env
232 simpl_binds env (bind:binds) = do { env' <- simpl_bind env bind
233 ; simpl_binds env' binds }
234
235 simpl_bind env (Rec pairs) = simplRecBind env TopLevel pairs
236 simpl_bind env (NonRec b r) = do { (env', b') <- addBndrRules env b (lookupRecBndr env b)
237 ; simplRecOrTopPair env' TopLevel NonRecursive b b' r }
238
239 {-
240 ************************************************************************
241 * *
242 \subsection{Lazy bindings}
243 * *
244 ************************************************************************
245
246 simplRecBind is used for
247 * recursive bindings only
248 -}
249
250 simplRecBind :: SimplEnv -> TopLevelFlag
251 -> [(InId, InExpr)]
252 -> SimplM SimplEnv
253 simplRecBind env0 top_lvl pairs0
254 = do { (env_with_info, triples) <- mapAccumLM add_rules env0 pairs0
255 ; env1 <- go (zapFloats env_with_info) triples
256 ; return (env0 `addRecFloats` env1) }
257 -- addFloats adds the floats from env1,
258 -- _and_ updates env0 with the in-scope set from env1
259 where
260 add_rules :: SimplEnv -> (InBndr,InExpr) -> SimplM (SimplEnv, (InBndr, OutBndr, InExpr))
261 -- Add the (substituted) rules to the binder
262 add_rules env (bndr, rhs)
263 = do { (env', bndr') <- addBndrRules env bndr (lookupRecBndr env bndr)
264 ; return (env', (bndr, bndr', rhs)) }
265
266 go env [] = return env
267
268 go env ((old_bndr, new_bndr, rhs) : pairs)
269 = do { env' <- simplRecOrTopPair env top_lvl Recursive old_bndr new_bndr rhs
270 ; go env' pairs }
271
272 {-
273 simplOrTopPair is used for
274 * recursive bindings (whether top level or not)
275 * top-level non-recursive bindings
276
277 It assumes the binder has already been simplified, but not its IdInfo.
278 -}
279
280 simplRecOrTopPair :: SimplEnv
281 -> TopLevelFlag -> RecFlag
282 -> InId -> OutBndr -> InExpr -- Binder and rhs
283 -> SimplM SimplEnv -- Returns an env that includes the binding
284
285 simplRecOrTopPair env top_lvl is_rec old_bndr new_bndr rhs
286 = do { dflags <- getDynFlags
287 ; trace_bind dflags $
288 if preInlineUnconditionally dflags env top_lvl old_bndr rhs
289 -- Check for unconditional inline
290 then do tick (PreInlineUnconditionally old_bndr)
291 return (extendIdSubst env old_bndr (mkContEx env rhs))
292 else simplLazyBind env top_lvl is_rec old_bndr new_bndr rhs env }
293 where
294 trace_bind dflags thing_inside
295 | not (dopt Opt_D_verbose_core2core dflags)
296 = thing_inside
297 | otherwise
298 = pprTrace "SimplBind" (ppr old_bndr) thing_inside
299 -- trace_bind emits a trace for each top-level binding, which
300 -- helps to locate the tracing for inlining and rule firing
301
302 {-
303 simplLazyBind is used for
304 * [simplRecOrTopPair] recursive bindings (whether top level or not)
305 * [simplRecOrTopPair] top-level non-recursive bindings
306 * [simplNonRecE] non-top-level *lazy* non-recursive bindings
307
308 Nota bene:
309 1. It assumes that the binder is *already* simplified,
310 and is in scope, and its IdInfo too, except unfolding
311
312 2. It assumes that the binder type is lifted.
313
314 3. It does not check for pre-inline-unconditionally;
315 that should have been done already.
316 -}
317
318 simplLazyBind :: SimplEnv
319 -> TopLevelFlag -> RecFlag
320 -> InId -> OutId -- Binder, both pre-and post simpl
321 -- The OutId has IdInfo, except arity, unfolding
322 -> InExpr -> SimplEnv -- The RHS and its environment
323 -> SimplM SimplEnv
324 -- Precondition: rhs obeys the let/app invariant
325 simplLazyBind env top_lvl is_rec bndr bndr1 rhs rhs_se
326 = -- pprTrace "simplLazyBind" ((ppr bndr <+> ppr bndr1) $$ ppr rhs $$ ppr (seIdSubst rhs_se)) $
327 do { let rhs_env = rhs_se `setInScope` env
328 (tvs, body) = case collectTyAndValBinders rhs of
329 (tvs, [], body)
330 | surely_not_lam body -> (tvs, body)
331 _ -> ([], rhs)
332
333 surely_not_lam (Lam {}) = False
334 surely_not_lam (Tick t e)
335 | not (tickishFloatable t) = surely_not_lam e
336 -- eta-reduction could float
337 surely_not_lam _ = True
338 -- Do not do the "abstract tyyvar" thing if there's
339 -- a lambda inside, because it defeats eta-reduction
340 -- f = /\a. \x. g a x
341 -- should eta-reduce.
342
343
344 ; (body_env, tvs') <- simplBinders rhs_env tvs
345 -- See Note [Floating and type abstraction] in SimplUtils
346
347 -- Simplify the RHS
348 ; let rhs_cont = mkRhsStop (substTy body_env (exprType body))
349 ; (body_env1, body1) <- simplExprF body_env body rhs_cont
350 -- ANF-ise a constructor or PAP rhs
351 ; (body_env2, body2) <- prepareRhs top_lvl body_env1 bndr1 body1
352
353 ; (env', rhs')
354 <- if not (doFloatFromRhs top_lvl is_rec False body2 body_env2)
355 then -- No floating, revert to body1
356 do { rhs' <- mkLam tvs' (wrapFloats body_env1 body1) rhs_cont
357 ; return (env, rhs') }
358
359 else if null tvs then -- Simple floating
360 do { tick LetFloatFromLet
361 ; return (addFloats env body_env2, body2) }
362
363 else -- Do type-abstraction first
364 do { tick LetFloatFromLet
365 ; (poly_binds, body3) <- abstractFloats tvs' body_env2 body2
366 ; rhs' <- mkLam tvs' body3 rhs_cont
367 ; env' <- foldlM (addPolyBind top_lvl) env poly_binds
368 ; return (env', rhs') }
369
370 ; completeBind env' top_lvl bndr bndr1 rhs' }
371
372 {-
373 A specialised variant of simplNonRec used when the RHS is already simplified,
374 notably in knownCon. It uses case-binding where necessary.
375 -}
376
377 simplNonRecX :: SimplEnv
378 -> InId -- Old binder
379 -> OutExpr -- Simplified RHS
380 -> SimplM SimplEnv
381 -- Precondition: rhs satisfies the let/app invariant
382 simplNonRecX env bndr new_rhs
383 | isDeadBinder bndr -- Not uncommon; e.g. case (a,b) of c { (p,q) -> p }
384 = return env -- Here c is dead, and we avoid creating
385 -- the binding c = (a,b)
386
387 | Coercion co <- new_rhs
388 = return (extendCvSubst env bndr co)
389
390 | otherwise
391 = do { (env', bndr') <- simplBinder env bndr
392 ; completeNonRecX NotTopLevel env' (isStrictId bndr) bndr bndr' new_rhs }
393 -- simplNonRecX is only used for NotTopLevel things
394
395 completeNonRecX :: TopLevelFlag -> SimplEnv
396 -> Bool
397 -> InId -- Old binder
398 -> OutId -- New binder
399 -> OutExpr -- Simplified RHS
400 -> SimplM SimplEnv
401 -- Precondition: rhs satisfies the let/app invariant
402 -- See Note [CoreSyn let/app invariant] in CoreSyn
403
404 completeNonRecX top_lvl env is_strict old_bndr new_bndr new_rhs
405 = do { (env1, rhs1) <- prepareRhs top_lvl (zapFloats env) new_bndr new_rhs
406 ; (env2, rhs2) <-
407 if doFloatFromRhs NotTopLevel NonRecursive is_strict rhs1 env1
408 then do { tick LetFloatFromLet
409 ; return (addFloats env env1, rhs1) } -- Add the floats to the main env
410 else return (env, wrapFloats env1 rhs1) -- Wrap the floats around the RHS
411 ; completeBind env2 NotTopLevel old_bndr new_bndr rhs2 }
412
413 {-
414 {- No, no, no! Do not try preInlineUnconditionally in completeNonRecX
415 Doing so risks exponential behaviour, because new_rhs has been simplified once already
416 In the cases described by the folowing commment, postInlineUnconditionally will
417 catch many of the relevant cases.
418 -- This happens; for example, the case_bndr during case of
419 -- known constructor: case (a,b) of x { (p,q) -> ... }
420 -- Here x isn't mentioned in the RHS, so we don't want to
421 -- create the (dead) let-binding let x = (a,b) in ...
422 --
423 -- Similarly, single occurrences can be inlined vigourously
424 -- e.g. case (f x, g y) of (a,b) -> ....
425 -- If a,b occur once we can avoid constructing the let binding for them.
426
427 Furthermore in the case-binding case preInlineUnconditionally risks extra thunks
428 -- Consider case I# (quotInt# x y) of
429 -- I# v -> let w = J# v in ...
430 -- If we gaily inline (quotInt# x y) for v, we end up building an
431 -- extra thunk:
432 -- let w = J# (quotInt# x y) in ...
433 -- because quotInt# can fail.
434
435 | preInlineUnconditionally env NotTopLevel bndr new_rhs
436 = thing_inside (extendIdSubst env bndr (DoneEx new_rhs))
437 -}
438
439 ----------------------------------
440 prepareRhs takes a putative RHS, checks whether it's a PAP or
441 constructor application and, if so, converts it to ANF, so that the
442 resulting thing can be inlined more easily. Thus
443 x = (f a, g b)
444 becomes
445 t1 = f a
446 t2 = g b
447 x = (t1,t2)
448
449 We also want to deal well cases like this
450 v = (f e1 `cast` co) e2
451 Here we want to make e1,e2 trivial and get
452 x1 = e1; x2 = e2; v = (f x1 `cast` co) v2
453 That's what the 'go' loop in prepareRhs does
454 -}
455
456 prepareRhs :: TopLevelFlag -> SimplEnv -> OutId -> OutExpr -> SimplM (SimplEnv, OutExpr)
457 -- Adds new floats to the env iff that allows us to return a good RHS
458 prepareRhs top_lvl env id (Cast rhs co) -- Note [Float coercions]
459 | Pair ty1 _ty2 <- coercionKind co -- Do *not* do this if rhs has an unlifted type
460 , not (isUnliftedType ty1) -- see Note [Float coercions (unlifted)]
461 = do { (env', rhs') <- makeTrivialWithInfo top_lvl env sanitised_info rhs
462 ; return (env', Cast rhs' co) }
463 where
464 sanitised_info = vanillaIdInfo `setStrictnessInfo` strictnessInfo info
465 `setDemandInfo` demandInfo info
466 info = idInfo id
467
468 prepareRhs top_lvl env0 _ rhs0
469 = do { (_is_exp, env1, rhs1) <- go 0 env0 rhs0
470 ; return (env1, rhs1) }
471 where
472 go n_val_args env (Cast rhs co)
473 = do { (is_exp, env', rhs') <- go n_val_args env rhs
474 ; return (is_exp, env', Cast rhs' co) }
475 go n_val_args env (App fun (Type ty))
476 = do { (is_exp, env', rhs') <- go n_val_args env fun
477 ; return (is_exp, env', App rhs' (Type ty)) }
478 go n_val_args env (App fun arg)
479 = do { (is_exp, env', fun') <- go (n_val_args+1) env fun
480 ; case is_exp of
481 True -> do { (env'', arg') <- makeTrivial top_lvl env' arg
482 ; return (True, env'', App fun' arg') }
483 False -> return (False, env, App fun arg) }
484 go n_val_args env (Var fun)
485 = return (is_exp, env, Var fun)
486 where
487 is_exp = isExpandableApp fun n_val_args -- The fun a constructor or PAP
488 -- See Note [CONLIKE pragma] in BasicTypes
489 -- The definition of is_exp should match that in
490 -- OccurAnal.occAnalApp
491
492 go n_val_args env (Tick t rhs)
493 -- We want to be able to float bindings past this
494 -- tick. Non-scoping ticks don't care.
495 | tickishScoped t == NoScope
496 = do { (is_exp, env', rhs') <- go n_val_args env rhs
497 ; return (is_exp, env', Tick t rhs') }
498 -- On the other hand, for scoping ticks we need to be able to
499 -- copy them on the floats, which in turn is only allowed if
500 -- we can obtain non-counting ticks.
501 | not (tickishCounts t) || tickishCanSplit t
502 = do { (is_exp, env', rhs') <- go n_val_args (zapFloats env) rhs
503 ; let tickIt (id, expr) = (id, mkTick (mkNoCount t) expr)
504 floats' = seFloats $ env `addFloats` mapFloats env' tickIt
505 ; return (is_exp, env' { seFloats = floats' }, Tick t rhs') }
506
507 go _ env other
508 = return (False, env, other)
509
510 {-
511 Note [Float coercions]
512 ~~~~~~~~~~~~~~~~~~~~~~
513 When we find the binding
514 x = e `cast` co
515 we'd like to transform it to
516 x' = e
517 x = x `cast` co -- A trivial binding
518 There's a chance that e will be a constructor application or function, or something
519 like that, so moving the coercion to the usage site may well cancel the coercions
520 and lead to further optimisation. Example:
521
522 data family T a :: *
523 data instance T Int = T Int
524
525 foo :: Int -> Int -> Int
526 foo m n = ...
527 where
528 x = T m
529 go 0 = 0
530 go n = case x of { T m -> go (n-m) }
531 -- This case should optimise
532
533 Note [Preserve strictness when floating coercions]
534 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
535 In the Note [Float coercions] transformation, keep the strictness info.
536 Eg
537 f = e `cast` co -- f has strictness SSL
538 When we transform to
539 f' = e -- f' also has strictness SSL
540 f = f' `cast` co -- f still has strictness SSL
541
542 Its not wrong to drop it on the floor, but better to keep it.
543
544 Note [Float coercions (unlifted)]
545 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
546 BUT don't do [Float coercions] if 'e' has an unlifted type.
547 This *can* happen:
548
549 foo :: Int = (error (# Int,Int #) "urk")
550 `cast` CoUnsafe (# Int,Int #) Int
551
552 If do the makeTrivial thing to the error call, we'll get
553 foo = case error (# Int,Int #) "urk" of v -> v `cast` ...
554 But 'v' isn't in scope!
555
556 These strange casts can happen as a result of case-of-case
557 bar = case (case x of { T -> (# 2,3 #); F -> error "urk" }) of
558 (# p,q #) -> p+q
559 -}
560
561 makeTrivialArg :: SimplEnv -> ArgSpec -> SimplM (SimplEnv, ArgSpec)
562 makeTrivialArg env (ValArg e) = do { (env', e') <- makeTrivial NotTopLevel env e
563 ; return (env', ValArg e') }
564 makeTrivialArg env arg = return (env, arg) -- CastBy, TyArg
565
566 makeTrivial :: TopLevelFlag -> SimplEnv -> OutExpr -> SimplM (SimplEnv, OutExpr)
567 -- Binds the expression to a variable, if it's not trivial, returning the variable
568 makeTrivial top_lvl env expr = makeTrivialWithInfo top_lvl env vanillaIdInfo expr
569
570 makeTrivialWithInfo :: TopLevelFlag -> SimplEnv -> IdInfo
571 -> OutExpr -> SimplM (SimplEnv, OutExpr)
572 -- Propagate strictness and demand info to the new binder
573 -- Note [Preserve strictness when floating coercions]
574 -- Returned SimplEnv has same substitution as incoming one
575 makeTrivialWithInfo top_lvl env info expr
576 | exprIsTrivial expr -- Already trivial
577 || not (bindingOk top_lvl expr expr_ty) -- Cannot trivialise
578 -- See Note [Cannot trivialise]
579 = return (env, expr)
580 | otherwise -- See Note [Take care] below
581 = do { uniq <- getUniqueM
582 ; let name = mkSystemVarName uniq (fsLit "a")
583 var = mkLocalIdOrCoVarWithInfo name expr_ty info
584 ; env' <- completeNonRecX top_lvl env False var var expr
585 ; expr' <- simplVar env' var
586 ; return (env', expr') }
587 -- The simplVar is needed becase we're constructing a new binding
588 -- a = rhs
589 -- And if rhs is of form (rhs1 |> co), then we might get
590 -- a1 = rhs1
591 -- a = a1 |> co
592 -- and now a's RHS is trivial and can be substituted out, and that
593 -- is what completeNonRecX will do
594 -- To put it another way, it's as if we'd simplified
595 -- let var = e in var
596 where
597 expr_ty = exprType expr
598
599 bindingOk :: TopLevelFlag -> CoreExpr -> Type -> Bool
600 -- True iff we can have a binding of this expression at this level
601 -- Precondition: the type is the type of the expression
602 bindingOk top_lvl _ expr_ty
603 | isTopLevel top_lvl = not (isUnliftedType expr_ty)
604 | otherwise = True
605
606 {-
607 Note [Cannot trivialise]
608 ~~~~~~~~~~~~~~~~~~~~~~~~
609 Consider tih
610 f :: Int -> Addr#
611
612 foo :: Bar
613 foo = Bar (f 3)
614
615 Then we can't ANF-ise foo, even though we'd like to, because
616 we can't make a top-level binding for the Addr# (f 3). And if
617 so we don't want to turn it into
618 foo = let x = f 3 in Bar x
619 because we'll just end up inlining x back, and that makes the
620 simplifier loop. Better not to ANF-ise it at all.
621
622 A case in point is literal strings (a MachStr is not regarded as
623 trivial):
624
625 foo = Ptr "blob"#
626
627 We don't want to ANF-ise this.
628
629 ************************************************************************
630 * *
631 \subsection{Completing a lazy binding}
632 * *
633 ************************************************************************
634
635 completeBind
636 * deals only with Ids, not TyVars
637 * takes an already-simplified binder and RHS
638 * is used for both recursive and non-recursive bindings
639 * is used for both top-level and non-top-level bindings
640
641 It does the following:
642 - tries discarding a dead binding
643 - tries PostInlineUnconditionally
644 - add unfolding [this is the only place we add an unfolding]
645 - add arity
646
647 It does *not* attempt to do let-to-case. Why? Because it is used for
648 - top-level bindings (when let-to-case is impossible)
649 - many situations where the "rhs" is known to be a WHNF
650 (so let-to-case is inappropriate).
651
652 Nor does it do the atomic-argument thing
653 -}
654
655 completeBind :: SimplEnv
656 -> TopLevelFlag -- Flag stuck into unfolding
657 -> InId -- Old binder
658 -> OutId -> OutExpr -- New binder and RHS
659 -> SimplM SimplEnv
660 -- completeBind may choose to do its work
661 -- * by extending the substitution (e.g. let x = y in ...)
662 -- * or by adding to the floats in the envt
663 --
664 -- Precondition: rhs obeys the let/app invariant
665 completeBind env top_lvl old_bndr new_bndr new_rhs
666 | isCoVar old_bndr
667 = case new_rhs of
668 Coercion co -> return (extendCvSubst env old_bndr co)
669 _ -> return (addNonRec env new_bndr new_rhs)
670
671 | otherwise
672 = ASSERT( isId new_bndr )
673 do { let old_info = idInfo old_bndr
674 old_unf = unfoldingInfo old_info
675 occ_info = occInfo old_info
676
677 -- Do eta-expansion on the RHS of the binding
678 -- See Note [Eta-expanding at let bindings] in SimplUtils
679 ; (new_arity, final_rhs) <- tryEtaExpandRhs env new_bndr new_rhs
680
681 -- Simplify the unfolding
682 ; new_unfolding <- simplLetUnfolding env top_lvl old_bndr final_rhs old_unf
683
684 ; dflags <- getDynFlags
685 ; if postInlineUnconditionally dflags env top_lvl new_bndr occ_info
686 final_rhs new_unfolding
687
688 -- Inline and discard the binding
689 then do { tick (PostInlineUnconditionally old_bndr)
690 ; return (extendIdSubst env old_bndr (DoneEx final_rhs)) }
691 -- Use the substitution to make quite, quite sure that the
692 -- substitution will happen, since we are going to discard the binding
693 else
694 do { let info1 = idInfo new_bndr `setArityInfo` new_arity
695
696 -- Unfolding info: Note [Setting the new unfolding]
697 info2 = info1 `setUnfoldingInfo` new_unfolding
698
699 -- Demand info: Note [Setting the demand info]
700 --
701 -- We also have to nuke demand info if for some reason
702 -- eta-expansion *reduces* the arity of the binding to less
703 -- than that of the strictness sig. This can happen: see Note [Arity decrease].
704 info3 | isEvaldUnfolding new_unfolding
705 || (case strictnessInfo info2 of
706 StrictSig dmd_ty -> new_arity < dmdTypeDepth dmd_ty)
707 = zapDemandInfo info2 `orElse` info2
708 | otherwise
709 = info2
710
711 final_id = new_bndr `setIdInfo` info3
712
713 ; -- pprTrace "Binding" (ppr final_id <+> ppr new_unfolding) $
714 return (addNonRec env final_id final_rhs) } }
715 -- The addNonRec adds it to the in-scope set too
716
717 ------------------------------
718 addPolyBind :: TopLevelFlag -> SimplEnv -> OutBind -> SimplM SimplEnv
719 -- Add a new binding to the environment, complete with its unfolding
720 -- but *do not* do postInlineUnconditionally, because we have already
721 -- processed some of the scope of the binding
722 -- We still want the unfolding though. Consider
723 -- let
724 -- x = /\a. let y = ... in Just y
725 -- in body
726 -- Then we float the y-binding out (via abstractFloats and addPolyBind)
727 -- but 'x' may well then be inlined in 'body' in which case we'd like the
728 -- opportunity to inline 'y' too.
729 --
730 -- INVARIANT: the arity is correct on the incoming binders
731
732 addPolyBind top_lvl env (NonRec poly_id rhs)
733 = do { unfolding <- simplLetUnfolding env top_lvl poly_id rhs noUnfolding
734 -- Assumes that poly_id did not have an INLINE prag
735 -- which is perhaps wrong. ToDo: think about this
736 ; let final_id = setIdInfo poly_id $
737 idInfo poly_id `setUnfoldingInfo` unfolding
738
739 ; return (addNonRec env final_id rhs) }
740
741 addPolyBind _ env bind@(Rec _)
742 = return (extendFloats env bind)
743 -- Hack: letrecs are more awkward, so we extend "by steam"
744 -- without adding unfoldings etc. At worst this leads to
745 -- more simplifier iterations
746
747 {- Note [Arity decrease]
748 ~~~~~~~~~~~~~~~~~~~~~~~~
749 Generally speaking the arity of a binding should not decrease. But it *can*
750 legitimately happen because of RULES. Eg
751 f = g Int
752 where g has arity 2, will have arity 2. But if there's a rewrite rule
753 g Int --> h
754 where h has arity 1, then f's arity will decrease. Here's a real-life example,
755 which is in the output of Specialise:
756
757 Rec {
758 $dm {Arity 2} = \d.\x. op d
759 {-# RULES forall d. $dm Int d = $s$dm #-}
760
761 dInt = MkD .... opInt ...
762 opInt {Arity 1} = $dm dInt
763
764 $s$dm {Arity 0} = \x. op dInt }
765
766 Here opInt has arity 1; but when we apply the rule its arity drops to 0.
767 That's why Specialise goes to a little trouble to pin the right arity
768 on specialised functions too.
769
770 Note [Setting the demand info]
771 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
772 If the unfolding is a value, the demand info may
773 go pear-shaped, so we nuke it. Example:
774 let x = (a,b) in
775 case x of (p,q) -> h p q x
776 Here x is certainly demanded. But after we've nuked
777 the case, we'll get just
778 let x = (a,b) in h a b x
779 and now x is not demanded (I'm assuming h is lazy)
780 This really happens. Similarly
781 let f = \x -> e in ...f..f...
782 After inlining f at some of its call sites the original binding may
783 (for example) be no longer strictly demanded.
784 The solution here is a bit ad hoc...
785
786
787 ************************************************************************
788 * *
789 \subsection[Simplify-simplExpr]{The main function: simplExpr}
790 * *
791 ************************************************************************
792
793 The reason for this OutExprStuff stuff is that we want to float *after*
794 simplifying a RHS, not before. If we do so naively we get quadratic
795 behaviour as things float out.
796
797 To see why it's important to do it after, consider this (real) example:
798
799 let t = f x
800 in fst t
801 ==>
802 let t = let a = e1
803 b = e2
804 in (a,b)
805 in fst t
806 ==>
807 let a = e1
808 b = e2
809 t = (a,b)
810 in
811 a -- Can't inline a this round, cos it appears twice
812 ==>
813 e1
814
815 Each of the ==> steps is a round of simplification. We'd save a
816 whole round if we float first. This can cascade. Consider
817
818 let f = g d
819 in \x -> ...f...
820 ==>
821 let f = let d1 = ..d.. in \y -> e
822 in \x -> ...f...
823 ==>
824 let d1 = ..d..
825 in \x -> ...(\y ->e)...
826
827 Only in this second round can the \y be applied, and it
828 might do the same again.
829 -}
830
831 simplExpr :: SimplEnv -> CoreExpr -> SimplM CoreExpr
832 simplExpr env expr = simplExprC env expr (mkBoringStop expr_out_ty)
833 where
834 expr_out_ty :: OutType
835 expr_out_ty = substTy env (exprType expr)
836
837 simplExprC :: SimplEnv -> CoreExpr -> SimplCont -> SimplM CoreExpr
838 -- Simplify an expression, given a continuation
839 simplExprC env expr cont
840 = -- pprTrace "simplExprC" (ppr expr $$ ppr cont {- $$ ppr (seIdSubst env) -} $$ ppr (seFloats env) ) $
841 do { (env', expr') <- simplExprF (zapFloats env) expr cont
842 ; -- pprTrace "simplExprC ret" (ppr expr $$ ppr expr') $
843 -- pprTrace "simplExprC ret3" (ppr (seInScope env')) $
844 -- pprTrace "simplExprC ret4" (ppr (seFloats env')) $
845 return (wrapFloats env' expr') }
846
847 --------------------------------------------------
848 simplExprF :: SimplEnv -> InExpr -> SimplCont
849 -> SimplM (SimplEnv, OutExpr)
850
851 simplExprF env e cont
852 = {- pprTrace "simplExprF" (vcat
853 [ ppr e
854 , text "cont =" <+> ppr cont
855 , text "inscope =" <+> ppr (seInScope env)
856 , text "tvsubst =" <+> ppr (seTvSubst env)
857 , text "idsubst =" <+> ppr (seIdSubst env)
858 , text "cvsubst =" <+> ppr (seCvSubst env)
859 {- , ppr (seFloats env) -}
860 ]) $ -}
861 simplExprF1 env e cont
862
863 simplExprF1 :: SimplEnv -> InExpr -> SimplCont
864 -> SimplM (SimplEnv, OutExpr)
865 simplExprF1 env (Var v) cont = simplIdF env v cont
866 simplExprF1 env (Lit lit) cont = rebuild env (Lit lit) cont
867 simplExprF1 env (Tick t expr) cont = simplTick env t expr cont
868 simplExprF1 env (Cast body co) cont = simplCast env body co cont
869 simplExprF1 env (Coercion co) cont = simplCoercionF env co cont
870 simplExprF1 env (Type ty) cont = ASSERT( contIsRhsOrArg cont )
871 rebuild env (Type (substTy env ty)) cont
872
873 simplExprF1 env (App fun arg) cont
874 = simplExprF env fun $
875 case arg of
876 Type ty -> ApplyToTy { sc_arg_ty = substTy env ty
877 , sc_hole_ty = substTy env (exprType fun)
878 , sc_cont = cont }
879 _ -> ApplyToVal { sc_arg = arg, sc_env = env
880 , sc_dup = NoDup, sc_cont = cont }
881
882 simplExprF1 env expr@(Lam {}) cont
883 = simplLam env zapped_bndrs body cont
884 -- The main issue here is under-saturated lambdas
885 -- (\x1. \x2. e) arg1
886 -- Here x1 might have "occurs-once" occ-info, because occ-info
887 -- is computed assuming that a group of lambdas is applied
888 -- all at once. If there are too few args, we must zap the
889 -- occ-info, UNLESS the remaining binders are one-shot
890 where
891 (bndrs, body) = collectBinders expr
892 zapped_bndrs | need_to_zap = map zap bndrs
893 | otherwise = bndrs
894
895 need_to_zap = any zappable_bndr (drop n_args bndrs)
896 n_args = countArgs cont
897 -- NB: countArgs counts all the args (incl type args)
898 -- and likewise drop counts all binders (incl type lambdas)
899
900 zappable_bndr b = isId b && not (isOneShotBndr b)
901 zap b | isTyVar b = b
902 | otherwise = zapLamIdInfo b
903
904 simplExprF1 env (Case scrut bndr _ alts) cont
905 = simplExprF env scrut (Select { sc_dup = NoDup, sc_bndr = bndr
906 , sc_alts = alts
907 , sc_env = env, sc_cont = cont })
908
909 simplExprF1 env (Let (Rec pairs) body) cont
910 = do { env' <- simplRecBndrs env (map fst pairs)
911 -- NB: bndrs' don't have unfoldings or rules
912 -- We add them as we go down
913
914 ; env'' <- simplRecBind env' NotTopLevel pairs
915 ; simplExprF env'' body cont }
916
917 simplExprF1 env (Let (NonRec bndr rhs) body) cont
918 = simplNonRecE env bndr (rhs, env) ([], body) cont
919
920 ---------------------------------
921 simplType :: SimplEnv -> InType -> SimplM OutType
922 -- Kept monadic just so we can do the seqType
923 simplType env ty
924 = -- pprTrace "simplType" (ppr ty $$ ppr (seTvSubst env)) $
925 seqType new_ty `seq` return new_ty
926 where
927 new_ty = substTy env ty
928
929 ---------------------------------
930 simplCoercionF :: SimplEnv -> InCoercion -> SimplCont
931 -> SimplM (SimplEnv, OutExpr)
932 simplCoercionF env co cont
933 = do { co' <- simplCoercion env co
934 ; rebuild env (Coercion co') cont }
935
936 simplCoercion :: SimplEnv -> InCoercion -> SimplM OutCoercion
937 simplCoercion env co
938 = let opt_co = optCoercion (getTCvSubst env) co
939 in seqCo opt_co `seq` return opt_co
940
941 -----------------------------------
942 -- | Push a TickIt context outwards past applications and cases, as
943 -- long as this is a non-scoping tick, to let case and application
944 -- optimisations apply.
945
946 simplTick :: SimplEnv -> Tickish Id -> InExpr -> SimplCont
947 -> SimplM (SimplEnv, OutExpr)
948 simplTick env tickish expr cont
949 -- A scoped tick turns into a continuation, so that we can spot
950 -- (scc t (\x . e)) in simplLam and eliminate the scc. If we didn't do
951 -- it this way, then it would take two passes of the simplifier to
952 -- reduce ((scc t (\x . e)) e').
953 -- NB, don't do this with counting ticks, because if the expr is
954 -- bottom, then rebuildCall will discard the continuation.
955
956 -- XXX: we cannot do this, because the simplifier assumes that
957 -- the context can be pushed into a case with a single branch. e.g.
958 -- scc<f> case expensive of p -> e
959 -- becomes
960 -- case expensive of p -> scc<f> e
961 --
962 -- So I'm disabling this for now. It just means we will do more
963 -- simplifier iterations that necessary in some cases.
964
965 -- | tickishScoped tickish && not (tickishCounts tickish)
966 -- = simplExprF env expr (TickIt tickish cont)
967
968 -- For unscoped or soft-scoped ticks, we are allowed to float in new
969 -- cost, so we simply push the continuation inside the tick. This
970 -- has the effect of moving the tick to the outside of a case or
971 -- application context, allowing the normal case and application
972 -- optimisations to fire.
973 | tickish `tickishScopesLike` SoftScope
974 = do { (env', expr') <- simplExprF env expr cont
975 ; return (env', mkTick tickish expr')
976 }
977
978 -- Push tick inside if the context looks like this will allow us to
979 -- do a case-of-case - see Note [case-of-scc-of-case]
980 | Select {} <- cont, Just expr' <- push_tick_inside
981 = simplExprF env expr' cont
982
983 -- We don't want to move the tick, but we might still want to allow
984 -- floats to pass through with appropriate wrapping (or not, see
985 -- wrap_floats below)
986 --- | not (tickishCounts tickish) || tickishCanSplit tickish
987 -- = wrap_floats
988
989 | otherwise
990 = no_floating_past_tick
991
992 where
993
994 -- Try to push tick inside a case, see Note [case-of-scc-of-case].
995 push_tick_inside =
996 case expr0 of
997 Case scrut bndr ty alts
998 -> Just $ Case (tickScrut scrut) bndr ty (map tickAlt alts)
999 _other -> Nothing
1000 where (ticks, expr0) = stripTicksTop movable (Tick tickish expr)
1001 movable t = not (tickishCounts t) ||
1002 t `tickishScopesLike` NoScope ||
1003 tickishCanSplit t
1004 tickScrut e = foldr mkTick e ticks
1005 -- Alternatives get annotated with all ticks that scope in some way,
1006 -- but we don't want to count entries.
1007 tickAlt (c,bs,e) = (c,bs, foldr mkTick e ts_scope)
1008 ts_scope = map mkNoCount $
1009 filter (not . (`tickishScopesLike` NoScope)) ticks
1010
1011 no_floating_past_tick =
1012 do { let (inc,outc) = splitCont cont
1013 ; (env', expr') <- simplExprF (zapFloats env) expr inc
1014 ; let tickish' = simplTickish env tickish
1015 ; (env'', expr'') <- rebuild (zapFloats env')
1016 (wrapFloats env' expr')
1017 (TickIt tickish' outc)
1018 ; return (addFloats env env'', expr'')
1019 }
1020
1021 -- Alternative version that wraps outgoing floats with the tick. This
1022 -- results in ticks being duplicated, as we don't make any attempt to
1023 -- eliminate the tick if we re-inline the binding (because the tick
1024 -- semantics allows unrestricted inlining of HNFs), so I'm not doing
1025 -- this any more. FloatOut will catch any real opportunities for
1026 -- floating.
1027 --
1028 -- wrap_floats =
1029 -- do { let (inc,outc) = splitCont cont
1030 -- ; (env', expr') <- simplExprF (zapFloats env) expr inc
1031 -- ; let tickish' = simplTickish env tickish
1032 -- ; let wrap_float (b,rhs) = (zapIdStrictness (setIdArity b 0),
1033 -- mkTick (mkNoCount tickish') rhs)
1034 -- -- when wrapping a float with mkTick, we better zap the Id's
1035 -- -- strictness info and arity, because it might be wrong now.
1036 -- ; let env'' = addFloats env (mapFloats env' wrap_float)
1037 -- ; rebuild env'' expr' (TickIt tickish' outc)
1038 -- }
1039
1040
1041 simplTickish env tickish
1042 | Breakpoint n ids <- tickish
1043 = Breakpoint n (map (getDoneId . substId env) ids)
1044 | otherwise = tickish
1045
1046 -- Push type application and coercion inside a tick
1047 splitCont :: SimplCont -> (SimplCont, SimplCont)
1048 splitCont cont@(ApplyToTy { sc_cont = tail }) = (cont { sc_cont = inc }, outc)
1049 where (inc,outc) = splitCont tail
1050 splitCont (CastIt co c) = (CastIt co inc, outc)
1051 where (inc,outc) = splitCont c
1052 splitCont other = (mkBoringStop (contHoleType other), other)
1053
1054 getDoneId (DoneId id) = id
1055 getDoneId (DoneEx e) = getIdFromTrivialExpr e -- Note [substTickish] in CoreSubst
1056 getDoneId other = pprPanic "getDoneId" (ppr other)
1057
1058 -- Note [case-of-scc-of-case]
1059 -- It's pretty important to be able to transform case-of-case when
1060 -- there's an SCC in the way. For example, the following comes up
1061 -- in nofib/real/compress/Encode.hs:
1062 --
1063 -- case scctick<code_string.r1>
1064 -- case $wcode_string_r13s wild_XC w1_s137 w2_s138 l_aje
1065 -- of _ { (# ww1_s13f, ww2_s13g, ww3_s13h #) ->
1066 -- (ww1_s13f, ww2_s13g, ww3_s13h)
1067 -- }
1068 -- of _ { (ww_s12Y, ww1_s12Z, ww2_s130) ->
1069 -- tick<code_string.f1>
1070 -- (ww_s12Y,
1071 -- ww1_s12Z,
1072 -- PTTrees.PT
1073 -- @ GHC.Types.Char @ GHC.Types.Int wild2_Xj ww2_s130 r_ajf)
1074 -- }
1075 --
1076 -- We really want this case-of-case to fire, because then the 3-tuple
1077 -- will go away (indeed, the CPR optimisation is relying on this
1078 -- happening). But the scctick is in the way - we need to push it
1079 -- inside to expose the case-of-case. So we perform this
1080 -- transformation on the inner case:
1081 --
1082 -- scctick c (case e of { p1 -> e1; ...; pn -> en })
1083 -- ==>
1084 -- case (scctick c e) of { p1 -> scc c e1; ...; pn -> scc c en }
1085 --
1086 -- So we've moved a constant amount of work out of the scc to expose
1087 -- the case. We only do this when the continuation is interesting: in
1088 -- for now, it has to be another Case (maybe generalise this later).
1089
1090 {-
1091 ************************************************************************
1092 * *
1093 \subsection{The main rebuilder}
1094 * *
1095 ************************************************************************
1096 -}
1097
1098 rebuild :: SimplEnv -> OutExpr -> SimplCont -> SimplM (SimplEnv, OutExpr)
1099 -- At this point the substitution in the SimplEnv should be irrelevant
1100 -- only the in-scope set and floats should matter
1101 rebuild env expr cont
1102 = case cont of
1103 Stop {} -> return (env, expr)
1104 TickIt t cont -> rebuild env (mkTick t expr) cont
1105 CastIt co cont -> rebuild env (mkCast expr co) cont
1106 -- NB: mkCast implements the (Coercion co |> g) optimisation
1107
1108 Select { sc_bndr = bndr, sc_alts = alts, sc_env = se, sc_cont = cont }
1109 -> rebuildCase (se `setFloats` env) expr bndr alts cont
1110
1111 StrictArg info _ cont -> rebuildCall env (info `addValArgTo` expr) cont
1112 StrictBind b bs body se cont -> do { env' <- simplNonRecX (se `setFloats` env) b expr
1113 -- expr satisfies let/app since it started life
1114 -- in a call to simplNonRecE
1115 ; simplLam env' bs body cont }
1116
1117 ApplyToTy { sc_arg_ty = ty, sc_cont = cont}
1118 -> rebuild env (App expr (Type ty)) cont
1119 ApplyToVal { sc_arg = arg, sc_env = se, sc_dup = dup_flag, sc_cont = cont}
1120 -- See Note [Avoid redundant simplification]
1121 | isSimplified dup_flag -> rebuild env (App expr arg) cont
1122 | otherwise -> do { arg' <- simplExpr (se `setInScope` env) arg
1123 ; rebuild env (App expr arg') cont }
1124
1125
1126 {-
1127 ************************************************************************
1128 * *
1129 \subsection{Lambdas}
1130 * *
1131 ************************************************************************
1132 -}
1133
1134 simplCast :: SimplEnv -> InExpr -> Coercion -> SimplCont
1135 -> SimplM (SimplEnv, OutExpr)
1136 simplCast env body co0 cont0
1137 = do { co1 <- simplCoercion env co0
1138 ; cont1 <- addCoerce co1 cont0
1139 ; simplExprF env body cont1 }
1140 where
1141 addCoerce co cont = add_coerce co (coercionKind co) cont
1142
1143 add_coerce _co (Pair s1 k1) cont -- co :: ty~ty
1144 | s1 `eqType` k1 = return cont -- is a no-op
1145
1146 add_coerce co1 (Pair s1 _k2) (CastIt co2 cont)
1147 | (Pair _l1 t1) <- coercionKind co2
1148 -- e |> (g1 :: S1~L) |> (g2 :: L~T1)
1149 -- ==>
1150 -- e, if S1=T1
1151 -- e |> (g1 . g2 :: S1~T1) otherwise
1152 --
1153 -- For example, in the initial form of a worker
1154 -- we may find (coerce T (coerce S (\x.e))) y
1155 -- and we'd like it to simplify to e[y/x] in one round
1156 -- of simplification
1157 , s1 `eqType` t1 = return cont -- The coerces cancel out
1158 | otherwise = return (CastIt (mkTransCo co1 co2) cont)
1159
1160 add_coerce co (Pair s1s2 _t1t2) cont@(ApplyToTy { sc_arg_ty = arg_ty, sc_cont = tail })
1161 -- (f |> g) ty ---> (f ty) |> (g @ ty)
1162 -- This implements the PushT rule from the paper
1163 | isForAllTy s1s2
1164 = do { cont' <- addCoerce new_cast tail
1165 ; return (cont { sc_cont = cont' }) }
1166 where
1167 new_cast = mkInstCo co (mkNomReflCo arg_ty)
1168
1169 add_coerce co (Pair s1s2 t1t2) (ApplyToVal { sc_arg = arg, sc_env = arg_se
1170 , sc_dup = dup, sc_cont = cont })
1171 | isFunTy s1s2 -- This implements the Push rule from the paper
1172 , isFunTy t1t2 -- Check t1t2 to ensure 'arg' is a value arg
1173 -- (e |> (g :: s1s2 ~ t1->t2)) f
1174 -- ===>
1175 -- (e (f |> (arg g :: t1~s1))
1176 -- |> (res g :: s2->t2)
1177 --
1178 -- t1t2 must be a function type, t1->t2, because it's applied
1179 -- to something but s1s2 might conceivably not be
1180 --
1181 -- When we build the ApplyTo we can't mix the out-types
1182 -- with the InExpr in the argument, so we simply substitute
1183 -- to make it all consistent. It's a bit messy.
1184 -- But it isn't a common case.
1185 --
1186 -- Example of use: Trac #995
1187 = do { (dup', arg_se', arg') <- simplArg env dup arg_se arg
1188 ; cont' <- addCoerce co2 cont
1189 ; return (ApplyToVal { sc_arg = mkCast arg' (mkSymCo co1)
1190 , sc_env = arg_se'
1191 , sc_dup = dup'
1192 , sc_cont = cont' }) }
1193 where
1194 -- we split coercion t1->t2 ~ s1->s2 into t1 ~ s1 and
1195 -- t2 ~ s2 with left and right on the curried form:
1196 -- (->) t1 t2 ~ (->) s1 s2
1197 [co1, co2] = decomposeCo 2 co
1198
1199 add_coerce co _ cont = return (CastIt co cont)
1200
1201 simplArg :: SimplEnv -> DupFlag -> StaticEnv -> CoreExpr
1202 -> SimplM (DupFlag, StaticEnv, OutExpr)
1203 simplArg env dup_flag arg_env arg
1204 | isSimplified dup_flag
1205 = return (dup_flag, arg_env, arg)
1206 | otherwise
1207 = do { arg' <- simplExpr (arg_env `setInScope` env) arg
1208 ; return (Simplified, zapSubstEnv arg_env, arg') }
1209
1210 {-
1211 ************************************************************************
1212 * *
1213 \subsection{Lambdas}
1214 * *
1215 ************************************************************************
1216
1217 Note [Zap unfolding when beta-reducing]
1218 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1219 Lambda-bound variables can have stable unfoldings, such as
1220 $j = \x. \b{Unf=Just x}. e
1221 See Note [Case binders and join points] below; the unfolding for lets
1222 us optimise e better. However when we beta-reduce it we want to
1223 revert to using the actual value, otherwise we can end up in the
1224 stupid situation of
1225 let x = blah in
1226 let b{Unf=Just x} = y
1227 in ...b...
1228 Here it'd be far better to drop the unfolding and use the actual RHS.
1229 -}
1230
1231 simplLam :: SimplEnv -> [InId] -> InExpr -> SimplCont
1232 -> SimplM (SimplEnv, OutExpr)
1233
1234 simplLam env [] body cont = simplExprF env body cont
1235
1236 -- Beta reduction
1237
1238 simplLam env (bndr:bndrs) body (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1239 = do { tick (BetaReduction bndr)
1240 ; simplLam (extendTvSubst env bndr arg_ty) bndrs body cont }
1241
1242 simplLam env (bndr:bndrs) body (ApplyToVal { sc_arg = arg, sc_env = arg_se
1243 , sc_cont = cont })
1244 = do { tick (BetaReduction bndr)
1245 ; simplNonRecE env' (zap_unfolding bndr) (arg, arg_se) (bndrs, body) cont }
1246 where
1247 env' | Coercion co <- arg
1248 = extendCvSubst env bndr co
1249 | otherwise
1250 = env
1251
1252 zap_unfolding bndr -- See Note [Zap unfolding when beta-reducing]
1253 | isId bndr, isStableUnfolding (realIdUnfolding bndr)
1254 = setIdUnfolding bndr NoUnfolding
1255 | otherwise = bndr
1256
1257 -- discard a non-counting tick on a lambda. This may change the
1258 -- cost attribution slightly (moving the allocation of the
1259 -- lambda elsewhere), but we don't care: optimisation changes
1260 -- cost attribution all the time.
1261 simplLam env bndrs body (TickIt tickish cont)
1262 | not (tickishCounts tickish)
1263 = simplLam env bndrs body cont
1264
1265 -- Not enough args, so there are real lambdas left to put in the result
1266 simplLam env bndrs body cont
1267 = do { (env', bndrs') <- simplLamBndrs env bndrs
1268 ; body' <- simplExpr env' body
1269 ; new_lam <- mkLam bndrs' body' cont
1270 ; rebuild env' new_lam cont }
1271
1272 simplLamBndrs :: SimplEnv -> [InBndr] -> SimplM (SimplEnv, [OutBndr])
1273 simplLamBndrs env bndrs = mapAccumLM simplLamBndr env bndrs
1274
1275 -------------
1276 simplLamBndr :: SimplEnv -> Var -> SimplM (SimplEnv, Var)
1277 -- Used for lambda binders. These sometimes have unfoldings added by
1278 -- the worker/wrapper pass that must be preserved, because they can't
1279 -- be reconstructed from context. For example:
1280 -- f x = case x of (a,b) -> fw a b x
1281 -- fw a b x{=(a,b)} = ...
1282 -- The "{=(a,b)}" is an unfolding we can't reconstruct otherwise.
1283 simplLamBndr env bndr
1284 | isId bndr && hasSomeUnfolding old_unf -- Special case
1285 = do { (env1, bndr1) <- simplBinder env bndr
1286 ; unf' <- simplUnfolding env1 NotTopLevel bndr old_unf
1287 ; let bndr2 = bndr1 `setIdUnfolding` unf'
1288 ; return (modifyInScope env1 bndr2, bndr2) }
1289
1290 | otherwise
1291 = simplBinder env bndr -- Normal case
1292 where
1293 old_unf = idUnfolding bndr
1294
1295 ------------------
1296 simplNonRecE :: SimplEnv
1297 -> InBndr -- The binder
1298 -> (InExpr, SimplEnv) -- Rhs of binding (or arg of lambda)
1299 -> ([InBndr], InExpr) -- Body of the let/lambda
1300 -- \xs.e
1301 -> SimplCont
1302 -> SimplM (SimplEnv, OutExpr)
1303
1304 -- simplNonRecE is used for
1305 -- * non-top-level non-recursive lets in expressions
1306 -- * beta reduction
1307 --
1308 -- It deals with strict bindings, via the StrictBind continuation,
1309 -- which may abort the whole process
1310 --
1311 -- Precondition: rhs satisfies the let/app invariant
1312 -- Note [CoreSyn let/app invariant] in CoreSyn
1313 --
1314 -- The "body" of the binding comes as a pair of ([InId],InExpr)
1315 -- representing a lambda; so we recurse back to simplLam
1316 -- Why? Because of the binder-occ-info-zapping done before
1317 -- the call to simplLam in simplExprF (Lam ...)
1318
1319 -- First deal with type applications and type lets
1320 -- (/\a. e) (Type ty) and (let a = Type ty in e)
1321 simplNonRecE env bndr (Type ty_arg, rhs_se) (bndrs, body) cont
1322 = ASSERT( isTyVar bndr )
1323 do { ty_arg' <- simplType (rhs_se `setInScope` env) ty_arg
1324 ; simplLam (extendTvSubst env bndr ty_arg') bndrs body cont }
1325
1326 simplNonRecE env bndr (rhs, rhs_se) (bndrs, body) cont
1327 = do dflags <- getDynFlags
1328 case () of
1329 _ | preInlineUnconditionally dflags env NotTopLevel bndr rhs
1330 -> do { tick (PreInlineUnconditionally bndr)
1331 ; -- pprTrace "preInlineUncond" (ppr bndr <+> ppr rhs) $
1332 simplLam (extendIdSubst env bndr (mkContEx rhs_se rhs)) bndrs body cont }
1333
1334 | isStrictId bndr -- Includes coercions
1335 -> simplExprF (rhs_se `setFloats` env) rhs
1336 (StrictBind bndr bndrs body env cont)
1337
1338 | otherwise
1339 -> ASSERT( not (isTyVar bndr) )
1340 do { (env1, bndr1) <- simplNonRecBndr env bndr
1341 ; (env2, bndr2) <- addBndrRules env1 bndr bndr1
1342 ; env3 <- simplLazyBind env2 NotTopLevel NonRecursive bndr bndr2 rhs rhs_se
1343 ; simplLam env3 bndrs body cont }
1344
1345 {-
1346 ************************************************************************
1347 * *
1348 Variables
1349 * *
1350 ************************************************************************
1351 -}
1352
1353 simplVar :: SimplEnv -> InVar -> SimplM OutExpr
1354 -- Look up an InVar in the environment
1355 simplVar env var
1356 | isTyVar var = return (Type (substTyVar env var))
1357 | isCoVar var = return (Coercion (substCoVar env var))
1358 | otherwise
1359 = case substId env var of
1360 DoneId var1 -> return (Var var1)
1361 DoneEx e -> return e
1362 ContEx tvs cvs ids e -> simplExpr (setSubstEnv env tvs cvs ids) e
1363
1364 simplIdF :: SimplEnv -> InId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1365 simplIdF env var cont
1366 = case substId env var of
1367 DoneEx e -> simplExprF (zapSubstEnv env) e cont
1368 ContEx tvs cvs ids e -> simplExprF (setSubstEnv env tvs cvs ids) e cont
1369 DoneId var1 -> completeCall env var1 cont
1370 -- Note [zapSubstEnv]
1371 -- The template is already simplified, so don't re-substitute.
1372 -- This is VITAL. Consider
1373 -- let x = e in
1374 -- let y = \z -> ...x... in
1375 -- \ x -> ...y...
1376 -- We'll clone the inner \x, adding x->x' in the id_subst
1377 -- Then when we inline y, we must *not* replace x by x' in
1378 -- the inlined copy!!
1379
1380 ---------------------------------------------------------
1381 -- Dealing with a call site
1382
1383 completeCall :: SimplEnv -> OutId -> SimplCont -> SimplM (SimplEnv, OutExpr)
1384 completeCall env var cont
1385 = do { ------------- Try inlining ----------------
1386 dflags <- getDynFlags
1387 ; let (lone_variable, arg_infos, call_cont) = contArgs cont
1388 n_val_args = length arg_infos
1389 interesting_cont = interestingCallContext call_cont
1390 unfolding = activeUnfolding env var
1391 maybe_inline = callSiteInline dflags var unfolding
1392 lone_variable arg_infos interesting_cont
1393 ; case maybe_inline of {
1394 Just expr -- There is an inlining!
1395 -> do { checkedTick (UnfoldingDone var)
1396 ; dump_inline dflags expr cont
1397 ; simplExprF (zapSubstEnv env) expr cont }
1398
1399 ; Nothing -> do -- No inlining!
1400
1401 { rule_base <- getSimplRules
1402 ; let info = mkArgInfo var (getRules rule_base var) n_val_args call_cont
1403 ; rebuildCall env info cont
1404 }}}
1405 where
1406 dump_inline dflags unfolding cont
1407 | not (dopt Opt_D_dump_inlinings dflags) = return ()
1408 | not (dopt Opt_D_verbose_core2core dflags)
1409 = when (isExternalName (idName var)) $
1410 liftIO $ printOutputForUser dflags alwaysQualify $
1411 sep [text "Inlining done:", nest 4 (ppr var)]
1412 | otherwise
1413 = liftIO $ printOutputForUser dflags alwaysQualify $
1414 sep [text "Inlining done: " <> ppr var,
1415 nest 4 (vcat [text "Inlined fn: " <+> nest 2 (ppr unfolding),
1416 text "Cont: " <+> ppr cont])]
1417
1418 rebuildCall :: SimplEnv
1419 -> ArgInfo
1420 -> SimplCont
1421 -> SimplM (SimplEnv, OutExpr)
1422 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_strs = [] }) cont
1423 -- When we run out of strictness args, it means
1424 -- that the call is definitely bottom; see SimplUtils.mkArgInfo
1425 -- Then we want to discard the entire strict continuation. E.g.
1426 -- * case (error "hello") of { ... }
1427 -- * (error "Hello") arg
1428 -- * f (error "Hello") where f is strict
1429 -- etc
1430 -- Then, especially in the first of these cases, we'd like to discard
1431 -- the continuation, leaving just the bottoming expression. But the
1432 -- type might not be right, so we may have to add a coerce.
1433 | not (contIsTrivial cont) -- Only do this if there is a non-trivial
1434 = return (env, castBottomExpr res cont_ty) -- contination to discard, else we do it
1435 where -- again and again!
1436 res = argInfoExpr fun rev_args
1437 cont_ty = contResultType cont
1438
1439 rebuildCall env info (CastIt co cont)
1440 = rebuildCall env (addCastTo info co) cont
1441
1442 rebuildCall env info (ApplyToTy { sc_arg_ty = arg_ty, sc_cont = cont })
1443 = rebuildCall env (info `addTyArgTo` arg_ty) cont
1444
1445 rebuildCall env info@(ArgInfo { ai_encl = encl_rules, ai_type = fun_ty
1446 , ai_strs = str:strs, ai_discs = disc:discs })
1447 (ApplyToVal { sc_arg = arg, sc_env = arg_se
1448 , sc_dup = dup_flag, sc_cont = cont })
1449 | isSimplified dup_flag -- See Note [Avoid redundant simplification]
1450 = rebuildCall env (addValArgTo info' arg) cont
1451
1452 | str -- Strict argument
1453 = -- pprTrace "Strict Arg" (ppr arg $$ ppr (seIdSubst env) $$ ppr (seInScope env)) $
1454 simplExprF (arg_se `setFloats` env) arg
1455 (StrictArg info' cci cont)
1456 -- Note [Shadowing]
1457
1458 | otherwise -- Lazy argument
1459 -- DO NOT float anything outside, hence simplExprC
1460 -- There is no benefit (unlike in a let-binding), and we'd
1461 -- have to be very careful about bogus strictness through
1462 -- floating a demanded let.
1463 = do { arg' <- simplExprC (arg_se `setInScope` env) arg
1464 (mkLazyArgStop (funArgTy fun_ty) cci)
1465 ; rebuildCall env (addValArgTo info' arg') cont }
1466 where
1467 info' = info { ai_strs = strs, ai_discs = discs }
1468 cci | encl_rules = RuleArgCtxt
1469 | disc > 0 = DiscArgCtxt -- Be keener here
1470 | otherwise = BoringCtxt -- Nothing interesting
1471
1472 rebuildCall env (ArgInfo { ai_fun = fun, ai_args = rev_args, ai_rules = rules }) cont
1473 | null rules
1474 = rebuild env (argInfoExpr fun rev_args) cont -- No rules, common case
1475
1476 | otherwise
1477 = do { -- We've accumulated a simplified call in <fun,rev_args>
1478 -- so try rewrite rules; see Note [RULEs apply to simplified arguments]
1479 -- See also Note [Rules for recursive functions]
1480 ; let env' = zapSubstEnv env -- See Note [zapSubstEnv];
1481 -- and NB that 'rev_args' are all fully simplified
1482 ; mb_rule <- tryRules env' rules fun (reverse rev_args) cont
1483 ; case mb_rule of {
1484 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1485
1486 -- Rules don't match
1487 ; Nothing -> rebuild env (argInfoExpr fun rev_args) cont -- No rules
1488 } }
1489
1490 {-
1491 Note [RULES apply to simplified arguments]
1492 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1493 It's very desirable to try RULES once the arguments have been simplified, because
1494 doing so ensures that rule cascades work in one pass. Consider
1495 {-# RULES g (h x) = k x
1496 f (k x) = x #-}
1497 ...f (g (h x))...
1498 Then we want to rewrite (g (h x)) to (k x) and only then try f's rules. If
1499 we match f's rules against the un-simplified RHS, it won't match. This
1500 makes a particularly big difference when superclass selectors are involved:
1501 op ($p1 ($p2 (df d)))
1502 We want all this to unravel in one sweep.
1503
1504 Note [Avoid redundant simplification]
1505 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1506 Because RULES apply to simplified arguments, there's a danger of repeatedly
1507 simplifying already-simplified arguments. An important example is that of
1508 (>>=) d e1 e2
1509 Here e1, e2 are simplified before the rule is applied, but don't really
1510 participate in the rule firing. So we mark them as Simplified to avoid
1511 re-simplifying them.
1512
1513 Note [Shadowing]
1514 ~~~~~~~~~~~~~~~~
1515 This part of the simplifier may break the no-shadowing invariant
1516 Consider
1517 f (...(\a -> e)...) (case y of (a,b) -> e')
1518 where f is strict in its second arg
1519 If we simplify the innermost one first we get (...(\a -> e)...)
1520 Simplifying the second arg makes us float the case out, so we end up with
1521 case y of (a,b) -> f (...(\a -> e)...) e'
1522 So the output does not have the no-shadowing invariant. However, there is
1523 no danger of getting name-capture, because when the first arg was simplified
1524 we used an in-scope set that at least mentioned all the variables free in its
1525 static environment, and that is enough.
1526
1527 We can't just do innermost first, or we'd end up with a dual problem:
1528 case x of (a,b) -> f e (...(\a -> e')...)
1529
1530 I spent hours trying to recover the no-shadowing invariant, but I just could
1531 not think of an elegant way to do it. The simplifier is already knee-deep in
1532 continuations. We have to keep the right in-scope set around; AND we have
1533 to get the effect that finding (error "foo") in a strict arg position will
1534 discard the entire application and replace it with (error "foo"). Getting
1535 all this at once is TOO HARD!
1536
1537
1538 ************************************************************************
1539 * *
1540 Rewrite rules
1541 * *
1542 ************************************************************************
1543 -}
1544
1545 tryRules :: SimplEnv -> [CoreRule]
1546 -> Id -> [ArgSpec] -> SimplCont
1547 -> SimplM (Maybe (CoreExpr, SimplCont))
1548 -- The SimplEnv already has zapSubstEnv applied to it
1549
1550 tryRules env rules fn args call_cont
1551 | null rules
1552 = return Nothing
1553 {- Disabled until we fix #8326
1554 | fn `hasKey` tagToEnumKey -- See Note [Optimising tagToEnum#]
1555 , [_type_arg, val_arg] <- args
1556 , Select dup bndr ((_,[],rhs1) : rest_alts) se cont <- call_cont
1557 , isDeadBinder bndr
1558 = do { dflags <- getDynFlags
1559 ; let enum_to_tag :: CoreAlt -> CoreAlt
1560 -- Takes K -> e into tagK# -> e
1561 -- where tagK# is the tag of constructor K
1562 enum_to_tag (DataAlt con, [], rhs)
1563 = ASSERT( isEnumerationTyCon (dataConTyCon con) )
1564 (LitAlt tag, [], rhs)
1565 where
1566 tag = mkMachInt dflags (toInteger (dataConTag con - fIRST_TAG))
1567 enum_to_tag alt = pprPanic "tryRules: tagToEnum" (ppr alt)
1568
1569 new_alts = (DEFAULT, [], rhs1) : map enum_to_tag rest_alts
1570 new_bndr = setIdType bndr intPrimTy
1571 -- The binder is dead, but should have the right type
1572 ; return (Just (val_arg, Select dup new_bndr new_alts se cont)) }
1573 -}
1574 | otherwise
1575 = do { dflags <- getDynFlags
1576 ; case lookupRule dflags (getUnfoldingInRuleMatch env) (activeRule env)
1577 fn (argInfoAppArgs args) rules of {
1578 Nothing -> return Nothing ; -- No rule matches
1579 Just (rule, rule_rhs) ->
1580 do { checkedTick (RuleFired (ru_name rule))
1581 ; let cont' = pushSimplifiedArgs env
1582 (drop (ruleArity rule) args)
1583 call_cont
1584 -- (ruleArity rule) says how many args the rule consumed
1585 ; dump dflags rule rule_rhs
1586 ; return (Just (rule_rhs, cont')) }}}
1587 where
1588 dump dflags rule rule_rhs
1589 | dopt Opt_D_dump_rule_rewrites dflags
1590 = log_rule dflags Opt_D_dump_rule_rewrites "Rule fired" $ vcat
1591 [ text "Rule:" <+> ftext (ru_name rule)
1592 , text "Before:" <+> hang (ppr fn) 2 (sep (map ppr args))
1593 , text "After: " <+> pprCoreExpr rule_rhs
1594 , text "Cont: " <+> ppr call_cont ]
1595
1596 | dopt Opt_D_dump_rule_firings dflags
1597 = log_rule dflags Opt_D_dump_rule_firings "Rule fired:" $
1598 ftext (ru_name rule)
1599
1600 | otherwise
1601 = return ()
1602
1603 log_rule dflags flag hdr details
1604 = liftIO . dumpSDoc dflags alwaysQualify flag "" $
1605 sep [text hdr, nest 4 details]
1606
1607 {-
1608 Note [Optimising tagToEnum#]
1609 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1610 If we have an enumeration data type:
1611
1612 data Foo = A | B | C
1613
1614 Then we want to transform
1615
1616 case tagToEnum# x of ==> case x of
1617 A -> e1 DEFAULT -> e1
1618 B -> e2 1# -> e2
1619 C -> e3 2# -> e3
1620
1621 thereby getting rid of the tagToEnum# altogether. If there was a DEFAULT
1622 alternative we retain it (remember it comes first). If not the case must
1623 be exhaustive, and we reflect that in the transformed version by adding
1624 a DEFAULT. Otherwise Lint complains that the new case is not exhaustive.
1625 See #8317.
1626
1627 Note [Rules for recursive functions]
1628 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1629 You might think that we shouldn't apply rules for a loop breaker:
1630 doing so might give rise to an infinite loop, because a RULE is
1631 rather like an extra equation for the function:
1632 RULE: f (g x) y = x+y
1633 Eqn: f a y = a-y
1634
1635 But it's too drastic to disable rules for loop breakers.
1636 Even the foldr/build rule would be disabled, because foldr
1637 is recursive, and hence a loop breaker:
1638 foldr k z (build g) = g k z
1639 So it's up to the programmer: rules can cause divergence
1640
1641
1642 ************************************************************************
1643 * *
1644 Rebuilding a case expression
1645 * *
1646 ************************************************************************
1647
1648 Note [Case elimination]
1649 ~~~~~~~~~~~~~~~~~~~~~~~
1650 The case-elimination transformation discards redundant case expressions.
1651 Start with a simple situation:
1652
1653 case x# of ===> let y# = x# in e
1654 y# -> e
1655
1656 (when x#, y# are of primitive type, of course). We can't (in general)
1657 do this for algebraic cases, because we might turn bottom into
1658 non-bottom!
1659
1660 The code in SimplUtils.prepareAlts has the effect of generalise this
1661 idea to look for a case where we're scrutinising a variable, and we
1662 know that only the default case can match. For example:
1663
1664 case x of
1665 0# -> ...
1666 DEFAULT -> ...(case x of
1667 0# -> ...
1668 DEFAULT -> ...) ...
1669
1670 Here the inner case is first trimmed to have only one alternative, the
1671 DEFAULT, after which it's an instance of the previous case. This
1672 really only shows up in eliminating error-checking code.
1673
1674 Note that SimplUtils.mkCase combines identical RHSs. So
1675
1676 case e of ===> case e of DEFAULT -> r
1677 True -> r
1678 False -> r
1679
1680 Now again the case may be elminated by the CaseElim transformation.
1681 This includes things like (==# a# b#)::Bool so that we simplify
1682 case ==# a# b# of { True -> x; False -> x }
1683 to just
1684 x
1685 This particular example shows up in default methods for
1686 comparison operations (e.g. in (>=) for Int.Int32)
1687
1688 Note [Case elimination: lifted case]
1689 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1690 If a case over a lifted type has a single alternative, and is being used
1691 as a strict 'let' (all isDeadBinder bndrs), we may want to do this
1692 transformation:
1693
1694 case e of r ===> let r = e in ...r...
1695 _ -> ...r...
1696
1697 (a) 'e' is already evaluated (it may so if e is a variable)
1698 Specifically we check (exprIsHNF e). In this case
1699 we can just allocate the WHNF directly with a let.
1700 or
1701 (b) 'x' is not used at all and e is ok-for-speculation
1702 The ok-for-spec bit checks that we don't lose any
1703 exceptions or divergence.
1704
1705 NB: it'd be *sound* to switch from case to let if the
1706 scrutinee was not yet WHNF but was guaranteed to
1707 converge; but sticking with case means we won't build a
1708 thunk
1709
1710 or
1711 (c) 'x' is used strictly in the body, and 'e' is a variable
1712 Then we can just substitute 'e' for 'x' in the body.
1713 See Note [Eliminating redundant seqs]
1714
1715 For (b), the "not used at all" test is important. Consider
1716 case (case a ># b of { True -> (p,q); False -> (q,p) }) of
1717 r -> blah
1718 The scrutinee is ok-for-speculation (it looks inside cases), but we do
1719 not want to transform to
1720 let r = case a ># b of { True -> (p,q); False -> (q,p) }
1721 in blah
1722 because that builds an unnecessary thunk.
1723
1724 Note [Eliminating redundant seqs]
1725 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1726 If we have this:
1727 case x of r { _ -> ..r.. }
1728 where 'r' is used strictly in (..r..), the case is effectively a 'seq'
1729 on 'x', but since 'r' is used strictly anyway, we can safely transform to
1730 (...x...)
1731
1732 Note that this can change the error behaviour. For example, we might
1733 transform
1734 case x of { _ -> error "bad" }
1735 --> error "bad"
1736 which is might be puzzling if 'x' currently lambda-bound, but later gets
1737 let-bound to (error "good").
1738
1739 Nevertheless, the paper "A semantics for imprecise exceptions" allows
1740 this transformation. If you want to fix the evaluation order, use
1741 'pseq'. See Trac #8900 for an example where the loss of this
1742 transformation bit us in practice.
1743
1744 See also Note [Empty case alternatives] in CoreSyn.
1745
1746 Just for reference, the original code (added Jan 13) looked like this:
1747 || case_bndr_evald_next rhs
1748
1749 case_bndr_evald_next :: CoreExpr -> Bool
1750 -- See Note [Case binder next]
1751 case_bndr_evald_next (Var v) = v == case_bndr
1752 case_bndr_evald_next (Cast e _) = case_bndr_evald_next e
1753 case_bndr_evald_next (App e _) = case_bndr_evald_next e
1754 case_bndr_evald_next (Case e _ _ _) = case_bndr_evald_next e
1755 case_bndr_evald_next _ = False
1756
1757 (This came up when fixing Trac #7542. See also Note [Eta reduction of
1758 an eval'd function] in CoreUtils.)
1759
1760
1761 Note [Case elimination: unlifted case]
1762 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1763 Consider
1764 case a +# b of r -> ...r...
1765 Then we do case-elimination (to make a let) followed by inlining,
1766 to get
1767 .....(a +# b)....
1768 If we have
1769 case indexArray# a i of r -> ...r...
1770 we might like to do the same, and inline the (indexArray# a i).
1771 But indexArray# is not okForSpeculation, so we don't build a let
1772 in rebuildCase (lest it get floated *out*), so the inlining doesn't
1773 happen either.
1774
1775 This really isn't a big deal I think. The let can be
1776
1777
1778 Further notes about case elimination
1779 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1780 Consider: test :: Integer -> IO ()
1781 test = print
1782
1783 Turns out that this compiles to:
1784 Print.test
1785 = \ eta :: Integer
1786 eta1 :: Void# ->
1787 case PrelNum.< eta PrelNum.zeroInteger of wild { __DEFAULT ->
1788 case hPutStr stdout
1789 (PrelNum.jtos eta ($w[] @ Char))
1790 eta1
1791 of wild1 { (# new_s, a4 #) -> PrelIO.lvl23 new_s }}
1792
1793 Notice the strange '<' which has no effect at all. This is a funny one.
1794 It started like this:
1795
1796 f x y = if x < 0 then jtos x
1797 else if y==0 then "" else jtos x
1798
1799 At a particular call site we have (f v 1). So we inline to get
1800
1801 if v < 0 then jtos x
1802 else if 1==0 then "" else jtos x
1803
1804 Now simplify the 1==0 conditional:
1805
1806 if v<0 then jtos v else jtos v
1807
1808 Now common-up the two branches of the case:
1809
1810 case (v<0) of DEFAULT -> jtos v
1811
1812 Why don't we drop the case? Because it's strict in v. It's technically
1813 wrong to drop even unnecessary evaluations, and in practice they
1814 may be a result of 'seq' so we *definitely* don't want to drop those.
1815 I don't really know how to improve this situation.
1816 -}
1817
1818 ---------------------------------------------------------
1819 -- Eliminate the case if possible
1820
1821 rebuildCase, reallyRebuildCase
1822 :: SimplEnv
1823 -> OutExpr -- Scrutinee
1824 -> InId -- Case binder
1825 -> [InAlt] -- Alternatives (inceasing order)
1826 -> SimplCont
1827 -> SimplM (SimplEnv, OutExpr)
1828
1829 --------------------------------------------------
1830 -- 1. Eliminate the case if there's a known constructor
1831 --------------------------------------------------
1832
1833 rebuildCase env scrut case_bndr alts cont
1834 | Lit lit <- scrut -- No need for same treatment as constructors
1835 -- because literals are inlined more vigorously
1836 , not (litIsLifted lit)
1837 = do { tick (KnownBranch case_bndr)
1838 ; case findAlt (LitAlt lit) alts of
1839 Nothing -> missingAlt env case_bndr alts cont
1840 Just (_, bs, rhs) -> simple_rhs bs rhs }
1841
1842 | Just (con, ty_args, other_args) <- exprIsConApp_maybe (getUnfoldingInRuleMatch env) scrut
1843 -- Works when the scrutinee is a variable with a known unfolding
1844 -- as well as when it's an explicit constructor application
1845 = do { tick (KnownBranch case_bndr)
1846 ; case findAlt (DataAlt con) alts of
1847 Nothing -> missingAlt env case_bndr alts cont
1848 Just (DEFAULT, bs, rhs) -> simple_rhs bs rhs
1849 Just (_, bs, rhs) -> knownCon env scrut con ty_args other_args
1850 case_bndr bs rhs cont
1851 }
1852 where
1853 simple_rhs bs rhs = ASSERT( null bs )
1854 do { env' <- simplNonRecX env case_bndr scrut
1855 -- scrut is a constructor application,
1856 -- hence satisfies let/app invariant
1857 ; simplExprF env' rhs cont }
1858
1859
1860 --------------------------------------------------
1861 -- 2. Eliminate the case if scrutinee is evaluated
1862 --------------------------------------------------
1863
1864 rebuildCase env scrut case_bndr alts@[(_, bndrs, rhs)] cont
1865 -- See if we can get rid of the case altogether
1866 -- See Note [Case elimination]
1867 -- mkCase made sure that if all the alternatives are equal,
1868 -- then there is now only one (DEFAULT) rhs
1869
1870 -- 2a. Dropping the case altogether, if
1871 -- a) it binds nothing (so it's really just a 'seq')
1872 -- b) evaluating the scrutinee has no side effects
1873 | is_plain_seq
1874 , exprOkForSideEffects scrut
1875 -- The entire case is dead, so we can drop it
1876 -- if the scrutinee converges without having imperative
1877 -- side effects or raising a Haskell exception
1878 -- See Note [PrimOp can_fail and has_side_effects] in PrimOp
1879 = simplExprF env rhs cont
1880
1881 -- 2b. Turn the case into a let, if
1882 -- a) it binds only the case-binder
1883 -- b) unlifted case: the scrutinee is ok-for-speculation
1884 -- lifted case: the scrutinee is in HNF (or will later be demanded)
1885 | all_dead_bndrs
1886 , if is_unlifted
1887 then exprOkForSpeculation scrut -- See Note [Case elimination: unlifted case]
1888 else exprIsHNF scrut -- See Note [Case elimination: lifted case]
1889 || scrut_is_demanded_var scrut
1890 = do { tick (CaseElim case_bndr)
1891 ; env' <- simplNonRecX env case_bndr scrut
1892 ; simplExprF env' rhs cont }
1893
1894 -- 2c. Try the seq rules if
1895 -- a) it binds only the case binder
1896 -- b) a rule for seq applies
1897 -- See Note [User-defined RULES for seq] in MkId
1898 | is_plain_seq
1899 = do { let scrut_ty = exprType scrut
1900 rhs_ty = substTy env (exprType rhs)
1901 out_args = [ TyArg { as_arg_ty = scrut_ty
1902 , as_hole_ty = seq_id_ty }
1903 , TyArg { as_arg_ty = rhs_ty
1904 , as_hole_ty = piResultTy seq_id_ty scrut_ty }
1905 , ValArg scrut]
1906 rule_cont = ApplyToVal { sc_dup = NoDup, sc_arg = rhs
1907 , sc_env = env, sc_cont = cont }
1908 env' = zapSubstEnv env
1909 -- Lazily evaluated, so we don't do most of this
1910
1911 ; rule_base <- getSimplRules
1912 ; mb_rule <- tryRules env' (getRules rule_base seqId) seqId out_args rule_cont
1913 ; case mb_rule of
1914 Just (rule_rhs, cont') -> simplExprF env' rule_rhs cont'
1915 Nothing -> reallyRebuildCase env scrut case_bndr alts cont }
1916 where
1917 is_unlifted = isUnliftedType (idType case_bndr)
1918 all_dead_bndrs = all isDeadBinder bndrs -- bndrs are [InId]
1919 is_plain_seq = all_dead_bndrs && isDeadBinder case_bndr -- Evaluation *only* for effect
1920 seq_id_ty = idType seqId
1921
1922 scrut_is_demanded_var :: CoreExpr -> Bool
1923 -- See Note [Eliminating redundant seqs]
1924 scrut_is_demanded_var (Cast s _) = scrut_is_demanded_var s
1925 scrut_is_demanded_var (Var _) = isStrictDmd (idDemandInfo case_bndr)
1926 scrut_is_demanded_var _ = False
1927
1928
1929 rebuildCase env scrut case_bndr alts cont
1930 = reallyRebuildCase env scrut case_bndr alts cont
1931
1932 --------------------------------------------------
1933 -- 3. Catch-all case
1934 --------------------------------------------------
1935
1936 reallyRebuildCase env scrut case_bndr alts cont
1937 = do { -- Prepare the continuation;
1938 -- The new subst_env is in place
1939 (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
1940
1941 -- Simplify the alternatives
1942 ; (scrut', case_bndr', alts') <- simplAlts env' scrut case_bndr alts dup_cont
1943
1944 ; dflags <- getDynFlags
1945 ; let alts_ty' = contResultType dup_cont
1946 ; case_expr <- mkCase dflags scrut' case_bndr' alts_ty' alts'
1947
1948 -- Notice that rebuild gets the in-scope set from env', not alt_env
1949 -- (which in any case is only build in simplAlts)
1950 -- The case binder *not* scope over the whole returned case-expression
1951 ; rebuild env' case_expr nodup_cont }
1952
1953 {-
1954 simplCaseBinder checks whether the scrutinee is a variable, v. If so,
1955 try to eliminate uses of v in the RHSs in favour of case_bndr; that
1956 way, there's a chance that v will now only be used once, and hence
1957 inlined.
1958
1959 Historical note: we use to do the "case binder swap" in the Simplifier
1960 so there were additional complications if the scrutinee was a variable.
1961 Now the binder-swap stuff is done in the occurrence analyer; see
1962 OccurAnal Note [Binder swap].
1963
1964 Note [knownCon occ info]
1965 ~~~~~~~~~~~~~~~~~~~~~~~~
1966 If the case binder is not dead, then neither are the pattern bound
1967 variables:
1968 case <any> of x { (a,b) ->
1969 case x of { (p,q) -> p } }
1970 Here (a,b) both look dead, but come alive after the inner case is eliminated.
1971 The point is that we bring into the envt a binding
1972 let x = (a,b)
1973 after the outer case, and that makes (a,b) alive. At least we do unless
1974 the case binder is guaranteed dead.
1975
1976 Note [Case alternative occ info]
1977 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1978 When we are simply reconstructing a case (the common case), we always
1979 zap the occurrence info on the binders in the alternatives. Even
1980 if the case binder is dead, the scrutinee is usually a variable, and *that*
1981 can bring the case-alternative binders back to life.
1982 See Note [Add unfolding for scrutinee]
1983
1984 Note [Improving seq]
1985 ~~~~~~~~~~~~~~~~~~~
1986 Consider
1987 type family F :: * -> *
1988 type instance F Int = Int
1989
1990 ... case e of x { DEFAULT -> rhs } ...
1991
1992 where x::F Int. Then we'd like to rewrite (F Int) to Int, getting
1993
1994 case e `cast` co of x'::Int
1995 I# x# -> let x = x' `cast` sym co
1996 in rhs
1997
1998 so that 'rhs' can take advantage of the form of x'.
1999
2000 Notice that Note [Case of cast] (in OccurAnal) may then apply to the result.
2001
2002 Nota Bene: We only do the [Improving seq] transformation if the
2003 case binder 'x' is actually used in the rhs; that is, if the case
2004 is *not* a *pure* seq.
2005 a) There is no point in adding the cast to a pure seq.
2006 b) There is a good reason not to: doing so would interfere
2007 with seq rules (Note [Built-in RULES for seq] in MkId).
2008 In particular, this [Improving seq] thing *adds* a cast
2009 while [Built-in RULES for seq] *removes* one, so they
2010 just flip-flop.
2011
2012 You might worry about
2013 case v of x { __DEFAULT ->
2014 ... case (v `cast` co) of y { I# -> ... }}
2015 This is a pure seq (since x is unused), so [Improving seq] won't happen.
2016 But it's ok: the simplifier will replace 'v' by 'x' in the rhs to get
2017 case v of x { __DEFAULT ->
2018 ... case (x `cast` co) of y { I# -> ... }}
2019 Now the outer case is not a pure seq, so [Improving seq] will happen,
2020 and then the inner case will disappear.
2021
2022 The need for [Improving seq] showed up in Roman's experiments. Example:
2023 foo :: F Int -> Int -> Int
2024 foo t n = t `seq` bar n
2025 where
2026 bar 0 = 0
2027 bar n = bar (n - case t of TI i -> i)
2028 Here we'd like to avoid repeated evaluating t inside the loop, by
2029 taking advantage of the `seq`.
2030
2031 At one point I did transformation in LiberateCase, but it's more
2032 robust here. (Otherwise, there's a danger that we'll simply drop the
2033 'seq' altogether, before LiberateCase gets to see it.)
2034 -}
2035
2036 simplAlts :: SimplEnv
2037 -> OutExpr
2038 -> InId -- Case binder
2039 -> [InAlt] -- Non-empty
2040 -> SimplCont
2041 -> SimplM (OutExpr, OutId, [OutAlt]) -- Includes the continuation
2042 -- Like simplExpr, this just returns the simplified alternatives;
2043 -- it does not return an environment
2044 -- The returned alternatives can be empty, none are possible
2045
2046 simplAlts env scrut case_bndr alts cont'
2047 = do { let env0 = zapFloats env
2048
2049 ; (env1, case_bndr1) <- simplBinder env0 case_bndr
2050
2051 ; fam_envs <- getFamEnvs
2052 ; (alt_env', scrut', case_bndr') <- improveSeq fam_envs env1 scrut
2053 case_bndr case_bndr1 alts
2054
2055 ; (imposs_deflt_cons, in_alts) <- prepareAlts scrut' case_bndr' alts
2056 -- NB: it's possible that the returned in_alts is empty: this is handled
2057 -- by the caller (rebuildCase) in the missingAlt function
2058
2059 ; alts' <- mapM (simplAlt alt_env' (Just scrut') imposs_deflt_cons case_bndr' cont') in_alts
2060 ; -- pprTrace "simplAlts" (ppr case_bndr $$ ppr alts_ty $$ ppr alts_ty' $$ ppr alts $$ ppr cont') $
2061 return (scrut', case_bndr', alts') }
2062
2063
2064 ------------------------------------
2065 improveSeq :: (FamInstEnv, FamInstEnv) -> SimplEnv
2066 -> OutExpr -> InId -> OutId -> [InAlt]
2067 -> SimplM (SimplEnv, OutExpr, OutId)
2068 -- Note [Improving seq]
2069 improveSeq fam_envs env scrut case_bndr case_bndr1 [(DEFAULT,_,_)]
2070 | not (isDeadBinder case_bndr) -- Not a pure seq! See Note [Improving seq]
2071 , Just (co, ty2) <- topNormaliseType_maybe fam_envs (idType case_bndr1)
2072 = do { case_bndr2 <- newId (fsLit "nt") ty2
2073 ; let rhs = DoneEx (Var case_bndr2 `Cast` mkSymCo co)
2074 env2 = extendIdSubst env case_bndr rhs
2075 ; return (env2, scrut `Cast` co, case_bndr2) }
2076
2077 improveSeq _ env scrut _ case_bndr1 _
2078 = return (env, scrut, case_bndr1)
2079
2080
2081 ------------------------------------
2082 simplAlt :: SimplEnv
2083 -> Maybe OutExpr -- The scrutinee
2084 -> [AltCon] -- These constructors can't be present when
2085 -- matching the DEFAULT alternative
2086 -> OutId -- The case binder
2087 -> SimplCont
2088 -> InAlt
2089 -> SimplM OutAlt
2090
2091 simplAlt env _ imposs_deflt_cons case_bndr' cont' (DEFAULT, bndrs, rhs)
2092 = ASSERT( null bndrs )
2093 do { let env' = addBinderUnfolding env case_bndr'
2094 (mkOtherCon imposs_deflt_cons)
2095 -- Record the constructors that the case-binder *can't* be.
2096 ; rhs' <- simplExprC env' rhs cont'
2097 ; return (DEFAULT, [], rhs') }
2098
2099 simplAlt env scrut' _ case_bndr' cont' (LitAlt lit, bndrs, rhs)
2100 = ASSERT( null bndrs )
2101 do { env' <- addAltUnfoldings env scrut' case_bndr' (Lit lit)
2102 ; rhs' <- simplExprC env' rhs cont'
2103 ; return (LitAlt lit, [], rhs') }
2104
2105 simplAlt env scrut' _ case_bndr' cont' (DataAlt con, vs, rhs)
2106 = do { -- Deal with the pattern-bound variables
2107 -- Mark the ones that are in ! positions in the
2108 -- data constructor as certainly-evaluated.
2109 -- NB: simplLamBinders preserves this eval info
2110 ; let vs_with_evals = add_evals (dataConRepStrictness con)
2111 ; (env', vs') <- simplLamBndrs env vs_with_evals
2112
2113 -- Bind the case-binder to (con args)
2114 ; let inst_tys' = tyConAppArgs (idType case_bndr')
2115 con_app :: OutExpr
2116 con_app = mkConApp2 con inst_tys' vs'
2117
2118 ; env'' <- addAltUnfoldings env' scrut' case_bndr' con_app
2119 ; rhs' <- simplExprC env'' rhs cont'
2120 ; return (DataAlt con, vs', rhs') }
2121 where
2122 -- add_evals records the evaluated-ness of the bound variables of
2123 -- a case pattern. This is *important*. Consider
2124 -- data T = T !Int !Int
2125 --
2126 -- case x of { T a b -> T (a+1) b }
2127 --
2128 -- We really must record that b is already evaluated so that we don't
2129 -- go and re-evaluate it when constructing the result.
2130 -- See Note [Data-con worker strictness] in MkId.hs
2131 add_evals the_strs
2132 = go vs the_strs
2133 where
2134 go [] [] = []
2135 go (v:vs') strs | isTyVar v = v : go vs' strs
2136 go (v:vs') (str:strs)
2137 | isMarkedStrict str = eval v : go vs' strs
2138 | otherwise = zap v : go vs' strs
2139 go _ _ = pprPanic "cat_evals"
2140 (ppr con $$
2141 ppr vs $$
2142 ppr_with_length the_strs $$
2143 ppr_with_length (dataConRepArgTys con) $$
2144 ppr_with_length (dataConRepStrictness con))
2145 where
2146 ppr_with_length list
2147 = ppr list <+> parens (text "length =" <+> ppr (length list))
2148 -- NB: If this panic triggers, note that
2149 -- NoStrictnessMark doesn't print!
2150
2151 zap v = zapIdOccInfo v -- See Note [Case alternative occ info]
2152 eval v = zap v `setIdUnfolding` evaldUnfolding
2153
2154 addAltUnfoldings :: SimplEnv -> Maybe OutExpr -> OutId -> OutExpr -> SimplM SimplEnv
2155 addAltUnfoldings env scrut case_bndr con_app
2156 = do { dflags <- getDynFlags
2157 ; let con_app_unf = mkSimpleUnfolding dflags con_app
2158 env1 = addBinderUnfolding env case_bndr con_app_unf
2159
2160 -- See Note [Add unfolding for scrutinee]
2161 env2 = case scrut of
2162 Just (Var v) -> addBinderUnfolding env1 v con_app_unf
2163 Just (Cast (Var v) co) -> addBinderUnfolding env1 v $
2164 mkSimpleUnfolding dflags (Cast con_app (mkSymCo co))
2165 _ -> env1
2166
2167 ; traceSmpl "addAltUnf" (vcat [ppr case_bndr <+> ppr scrut, ppr con_app])
2168 ; return env2 }
2169
2170 addBinderUnfolding :: SimplEnv -> Id -> Unfolding -> SimplEnv
2171 addBinderUnfolding env bndr unf
2172 | debugIsOn, Just tmpl <- maybeUnfoldingTemplate unf
2173 = WARN( not (eqType (idType bndr) (exprType tmpl)),
2174 ppr bndr $$ ppr (idType bndr) $$ ppr tmpl $$ ppr (exprType tmpl) )
2175 modifyInScope env (bndr `setIdUnfolding` unf)
2176
2177 | otherwise
2178 = modifyInScope env (bndr `setIdUnfolding` unf)
2179
2180 zapBndrOccInfo :: Bool -> Id -> Id
2181 -- Consider case e of b { (a,b) -> ... }
2182 -- Then if we bind b to (a,b) in "...", and b is not dead,
2183 -- then we must zap the deadness info on a,b
2184 zapBndrOccInfo keep_occ_info pat_id
2185 | keep_occ_info = pat_id
2186 | otherwise = zapIdOccInfo pat_id
2187
2188 {-
2189 Note [Add unfolding for scrutinee]
2190 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2191 In general it's unlikely that a variable scrutinee will appear
2192 in the case alternatives case x of { ...x unlikely to appear... }
2193 because the binder-swap in OccAnal has got rid of all such occcurrences
2194 See Note [Binder swap] in OccAnal.
2195
2196 BUT it is still VERY IMPORTANT to add a suitable unfolding for a
2197 variable scrutinee, in simplAlt. Here's why
2198 case x of y
2199 (a,b) -> case b of c
2200 I# v -> ...(f y)...
2201 There is no occurrence of 'b' in the (...(f y)...). But y gets
2202 the unfolding (a,b), and *that* mentions b. If f has a RULE
2203 RULE f (p, I# q) = ...
2204 we want that rule to match, so we must extend the in-scope env with a
2205 suitable unfolding for 'y'. It's *essential* for rule matching; but
2206 it's also good for case-elimintation -- suppose that 'f' was inlined
2207 and did multi-level case analysis, then we'd solve it in one
2208 simplifier sweep instead of two.
2209
2210 Exactly the same issue arises in SpecConstr;
2211 see Note [Add scrutinee to ValueEnv too] in SpecConstr
2212
2213 HOWEVER, given
2214 case x of y { Just a -> r1; Nothing -> r2 }
2215 we do not want to add the unfolding x -> y to 'x', which might seem cool,
2216 since 'y' itself has different unfoldings in r1 and r2. Reason: if we
2217 did that, we'd have to zap y's deadness info and that is a very useful
2218 piece of information.
2219
2220 So instead we add the unfolding x -> Just a, and x -> Nothing in the
2221 respective RHSs.
2222
2223
2224 ************************************************************************
2225 * *
2226 \subsection{Known constructor}
2227 * *
2228 ************************************************************************
2229
2230 We are a bit careful with occurrence info. Here's an example
2231
2232 (\x* -> case x of (a*, b) -> f a) (h v, e)
2233
2234 where the * means "occurs once". This effectively becomes
2235 case (h v, e) of (a*, b) -> f a)
2236 and then
2237 let a* = h v; b = e in f a
2238 and then
2239 f (h v)
2240
2241 All this should happen in one sweep.
2242 -}
2243
2244 knownCon :: SimplEnv
2245 -> OutExpr -- The scrutinee
2246 -> DataCon -> [OutType] -> [OutExpr] -- The scrutinee (in pieces)
2247 -> InId -> [InBndr] -> InExpr -- The alternative
2248 -> SimplCont
2249 -> SimplM (SimplEnv, OutExpr)
2250
2251 knownCon env scrut dc dc_ty_args dc_args bndr bs rhs cont
2252 = do { env' <- bind_args env bs dc_args
2253 ; env'' <- bind_case_bndr env'
2254 ; simplExprF env'' rhs cont }
2255 where
2256 zap_occ = zapBndrOccInfo (isDeadBinder bndr) -- bndr is an InId
2257
2258 -- Ugh!
2259 bind_args env' [] _ = return env'
2260
2261 bind_args env' (b:bs') (Type ty : args)
2262 = ASSERT( isTyVar b )
2263 bind_args (extendTvSubst env' b ty) bs' args
2264
2265 bind_args env' (b:bs') (Coercion co : args)
2266 = ASSERT( isCoVar b )
2267 bind_args (extendCvSubst env' b co) bs' args
2268
2269 bind_args env' (b:bs') (arg : args)
2270 = ASSERT( isId b )
2271 do { let b' = zap_occ b
2272 -- Note that the binder might be "dead", because it doesn't
2273 -- occur in the RHS; and simplNonRecX may therefore discard
2274 -- it via postInlineUnconditionally.
2275 -- Nevertheless we must keep it if the case-binder is alive,
2276 -- because it may be used in the con_app. See Note [knownCon occ info]
2277 ; env'' <- simplNonRecX env' b' arg -- arg satisfies let/app invariant
2278 ; bind_args env'' bs' args }
2279
2280 bind_args _ _ _ =
2281 pprPanic "bind_args" $ ppr dc $$ ppr bs $$ ppr dc_args $$
2282 text "scrut:" <+> ppr scrut
2283
2284 -- It's useful to bind bndr to scrut, rather than to a fresh
2285 -- binding x = Con arg1 .. argn
2286 -- because very often the scrut is a variable, so we avoid
2287 -- creating, and then subsequently eliminating, a let-binding
2288 -- BUT, if scrut is a not a variable, we must be careful
2289 -- about duplicating the arg redexes; in that case, make
2290 -- a new con-app from the args
2291 bind_case_bndr env
2292 | isDeadBinder bndr = return env
2293 | exprIsTrivial scrut = return (extendIdSubst env bndr (DoneEx scrut))
2294 | otherwise = do { dc_args <- mapM (simplVar env) bs
2295 -- dc_ty_args are aready OutTypes,
2296 -- but bs are InBndrs
2297 ; let con_app = Var (dataConWorkId dc)
2298 `mkTyApps` dc_ty_args
2299 `mkApps` dc_args
2300 ; simplNonRecX env bndr con_app }
2301
2302 -------------------
2303 missingAlt :: SimplEnv -> Id -> [InAlt] -> SimplCont -> SimplM (SimplEnv, OutExpr)
2304 -- This isn't strictly an error, although it is unusual.
2305 -- It's possible that the simplifer might "see" that
2306 -- an inner case has no accessible alternatives before
2307 -- it "sees" that the entire branch of an outer case is
2308 -- inaccessible. So we simply put an error case here instead.
2309 missingAlt env case_bndr _ cont
2310 = WARN( True, text "missingAlt" <+> ppr case_bndr )
2311 return (env, mkImpossibleExpr (contResultType cont))
2312
2313 {-
2314 ************************************************************************
2315 * *
2316 \subsection{Duplicating continuations}
2317 * *
2318 ************************************************************************
2319 -}
2320
2321 prepareCaseCont :: SimplEnv
2322 -> [InAlt] -> SimplCont
2323 -> SimplM (SimplEnv,
2324 SimplCont, -- Dupable part
2325 SimplCont) -- Non-dupable part
2326 -- We are considering
2327 -- K[case _ of { p1 -> r1; ...; pn -> rn }]
2328 -- where K is some enclosing continuation for the case
2329 -- Goal: split K into two pieces Kdup,Knodup so that
2330 -- a) Kdup can be duplicated
2331 -- b) Knodup[Kdup[e]] = K[e]
2332 -- The idea is that we'll transform thus:
2333 -- Knodup[ (case _ of { p1 -> Kdup[r1]; ...; pn -> Kdup[rn] }
2334 --
2335 -- We may also return some extra bindings in SimplEnv (that scope over
2336 -- the entire continuation)
2337 --
2338 -- When case-of-case is off, just make the entire continuation non-dupable
2339
2340 prepareCaseCont env alts cont
2341 | not (sm_case_case (getMode env)) = return (env, mkBoringStop (contHoleType cont), cont)
2342 | not (many_alts alts) = return (env, cont, mkBoringStop (contResultType cont))
2343 | otherwise = mkDupableCont env cont
2344 where
2345 many_alts :: [InAlt] -> Bool -- True iff strictly > 1 non-bottom alternative
2346 many_alts [] = False -- See Note [Bottom alternatives]
2347 many_alts [_] = False
2348 many_alts (alt:alts)
2349 | is_bot_alt alt = many_alts alts
2350 | otherwise = not (all is_bot_alt alts)
2351
2352 is_bot_alt (_,_,rhs) = exprIsBottom rhs
2353
2354 {-
2355 Note [Bottom alternatives]
2356 ~~~~~~~~~~~~~~~~~~~~~~~~~~
2357 When we have
2358 case (case x of { A -> error .. ; B -> e; C -> error ..)
2359 of alts
2360 then we can just duplicate those alts because the A and C cases
2361 will disappear immediately. This is more direct than creating
2362 join points and inlining them away; and in some cases we would
2363 not even create the join points (see Note [Single-alternative case])
2364 and we would keep the case-of-case which is silly. See Trac #4930.
2365 -}
2366
2367 mkDupableCont :: SimplEnv -> SimplCont
2368 -> SimplM (SimplEnv, SimplCont, SimplCont)
2369
2370 mkDupableCont env cont
2371 | contIsDupable cont
2372 = return (env, cont, mkBoringStop (contResultType cont))
2373
2374 mkDupableCont _ (Stop {}) = panic "mkDupableCont" -- Handled by previous eqn
2375
2376 mkDupableCont env (CastIt ty cont)
2377 = do { (env', dup, nodup) <- mkDupableCont env cont
2378 ; return (env', CastIt ty dup, nodup) }
2379
2380 -- Duplicating ticks for now, not sure if this is good or not
2381 mkDupableCont env cont@(TickIt{})
2382 = return (env, mkBoringStop (contHoleType cont), cont)
2383
2384 mkDupableCont env cont@(StrictBind {})
2385 = return (env, mkBoringStop (contHoleType cont), cont)
2386 -- See Note [Duplicating StrictBind]
2387
2388 mkDupableCont env (StrictArg info cci cont)
2389 -- See Note [Duplicating StrictArg]
2390 = do { (env', dup, nodup) <- mkDupableCont env cont
2391 ; (env'', args') <- mapAccumLM makeTrivialArg env' (ai_args info)
2392 ; return (env'', StrictArg (info { ai_args = args' }) cci dup, nodup) }
2393
2394 mkDupableCont env cont@(ApplyToTy { sc_cont = tail })
2395 = do { (env', dup_cont, nodup_cont) <- mkDupableCont env tail
2396 ; return (env', cont { sc_cont = dup_cont }, nodup_cont ) }
2397
2398 mkDupableCont env (ApplyToVal { sc_arg = arg, sc_dup = dup, sc_env = se, sc_cont = cont })
2399 = -- e.g. [...hole...] (...arg...)
2400 -- ==>
2401 -- let a = ...arg...
2402 -- in [...hole...] a
2403 do { (env', dup_cont, nodup_cont) <- mkDupableCont env cont
2404 ; (_, se', arg') <- simplArg env' dup se arg
2405 ; (env'', arg'') <- makeTrivial NotTopLevel env' arg'
2406 ; let app_cont = ApplyToVal { sc_arg = arg'', sc_env = se'
2407 , sc_dup = OkToDup, sc_cont = dup_cont }
2408 ; return (env'', app_cont, nodup_cont) }
2409
2410 mkDupableCont env cont@(Select { sc_bndr = case_bndr, sc_alts = [(_, bs, _rhs)] })
2411 -- See Note [Single-alternative case]
2412 -- | not (exprIsDupable rhs && contIsDupable case_cont)
2413 -- | not (isDeadBinder case_bndr)
2414 | all isDeadBinder bs -- InIds
2415 && not (isUnliftedType (idType case_bndr))
2416 -- Note [Single-alternative-unlifted]
2417 = return (env, mkBoringStop (contHoleType cont), cont)
2418
2419 mkDupableCont env (Select { sc_bndr = case_bndr, sc_alts = alts
2420 , sc_env = se, sc_cont = cont })
2421 = -- e.g. (case [...hole...] of { pi -> ei })
2422 -- ===>
2423 -- let ji = \xij -> ei
2424 -- in case [...hole...] of { pi -> ji xij }
2425 do { tick (CaseOfCase case_bndr)
2426 ; (env', dup_cont, nodup_cont) <- prepareCaseCont env alts cont
2427 -- NB: We call prepareCaseCont here. If there is only one
2428 -- alternative, then dup_cont may be big, but that's ok
2429 -- because we push it into the single alternative, and then
2430 -- use mkDupableAlt to turn that simplified alternative into
2431 -- a join point if it's too big to duplicate.
2432 -- And this is important: see Note [Fusing case continuations]
2433
2434 ; let alt_env = se `setInScope` env'
2435
2436 ; (alt_env', case_bndr') <- simplBinder alt_env case_bndr
2437 ; alts' <- mapM (simplAlt alt_env' Nothing [] case_bndr' dup_cont) alts
2438 -- Safe to say that there are no handled-cons for the DEFAULT case
2439 -- NB: simplBinder does not zap deadness occ-info, so
2440 -- a dead case_bndr' will still advertise its deadness
2441 -- This is really important because in
2442 -- case e of b { (# p,q #) -> ... }
2443 -- b is always dead, and indeed we are not allowed to bind b to (# p,q #),
2444 -- which might happen if e was an explicit unboxed pair and b wasn't marked dead.
2445 -- In the new alts we build, we have the new case binder, so it must retain
2446 -- its deadness.
2447 -- NB: we don't use alt_env further; it has the substEnv for
2448 -- the alternatives, and we don't want that
2449
2450 ; (env'', alts'') <- mkDupableAlts env' case_bndr' alts'
2451 ; return (env'', -- Note [Duplicated env]
2452 Select { sc_dup = OkToDup
2453 , sc_bndr = case_bndr', sc_alts = alts''
2454 , sc_env = zapSubstEnv env''
2455 , sc_cont = mkBoringStop (contHoleType nodup_cont) },
2456 nodup_cont) }
2457
2458
2459 mkDupableAlts :: SimplEnv -> OutId -> [InAlt]
2460 -> SimplM (SimplEnv, [InAlt])
2461 -- Absorbs the continuation into the new alternatives
2462
2463 mkDupableAlts env case_bndr' the_alts
2464 = go env the_alts
2465 where
2466 go env0 [] = return (env0, [])
2467 go env0 (alt:alts)
2468 = do { (env1, alt') <- mkDupableAlt env0 case_bndr' alt
2469 ; (env2, alts') <- go env1 alts
2470 ; return (env2, alt' : alts' ) }
2471
2472 mkDupableAlt :: SimplEnv -> OutId -> (AltCon, [CoreBndr], CoreExpr)
2473 -> SimplM (SimplEnv, (AltCon, [CoreBndr], CoreExpr))
2474 mkDupableAlt env case_bndr (con, bndrs', rhs') = do
2475 dflags <- getDynFlags
2476 if exprIsDupable dflags rhs' -- Note [Small alternative rhs]
2477 then return (env, (con, bndrs', rhs'))
2478 else
2479 do { let rhs_ty' = exprType rhs'
2480 scrut_ty = idType case_bndr
2481 case_bndr_w_unf
2482 = case con of
2483 DEFAULT -> case_bndr
2484 DataAlt dc -> setIdUnfolding case_bndr unf
2485 where
2486 -- See Note [Case binders and join points]
2487 unf = mkInlineUnfolding Nothing rhs
2488 rhs = mkConApp2 dc (tyConAppArgs scrut_ty) bndrs'
2489
2490 LitAlt {} -> WARN( True, text "mkDupableAlt"
2491 <+> ppr case_bndr <+> ppr con )
2492 case_bndr
2493 -- The case binder is alive but trivial, so why has
2494 -- it not been substituted away?
2495
2496 used_bndrs' | isDeadBinder case_bndr = filter abstract_over bndrs'
2497 | otherwise = bndrs' ++ [case_bndr_w_unf]
2498
2499 abstract_over bndr
2500 | isTyVar bndr = True -- Abstract over all type variables just in case
2501 | otherwise = not (isDeadBinder bndr)
2502 -- The deadness info on the new Ids is preserved by simplBinders
2503
2504 ; (final_bndrs', final_args) -- Note [Join point abstraction]
2505 <- if (any isId used_bndrs')
2506 then return (used_bndrs', varsToCoreExprs used_bndrs')
2507 else do { rw_id <- newId (fsLit "w") voidPrimTy
2508 ; return ([setOneShotLambda rw_id], [Var voidPrimId]) }
2509
2510 ; join_bndr <- newId (fsLit "$j") (mkPiTypes final_bndrs' rhs_ty')
2511 -- Note [Funky mkPiTypes]
2512
2513 ; let -- We make the lambdas into one-shot-lambdas. The
2514 -- join point is sure to be applied at most once, and doing so
2515 -- prevents the body of the join point being floated out by
2516 -- the full laziness pass
2517 really_final_bndrs = map one_shot final_bndrs'
2518 one_shot v | isId v = setOneShotLambda v
2519 | otherwise = v
2520 join_rhs = mkLams really_final_bndrs rhs'
2521 join_arity = exprArity join_rhs
2522 join_call = mkApps (Var join_bndr) final_args
2523
2524 ; env' <- addPolyBind NotTopLevel env (NonRec (join_bndr `setIdArity` join_arity) join_rhs)
2525 ; return (env', (con, bndrs', join_call)) }
2526 -- See Note [Duplicated env]
2527
2528 {-
2529 Note [Fusing case continuations]
2530 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2531 It's important to fuse two successive case continuations when the
2532 first has one alternative. That's why we call prepareCaseCont here.
2533 Consider this, which arises from thunk splitting (see Note [Thunk
2534 splitting] in WorkWrap):
2535
2536 let
2537 x* = case (case v of {pn -> rn}) of
2538 I# a -> I# a
2539 in body
2540
2541 The simplifier will find
2542 (Var v) with continuation
2543 Select (pn -> rn) (
2544 Select [I# a -> I# a] (
2545 StrictBind body Stop
2546
2547 So we'll call mkDupableCont on
2548 Select [I# a -> I# a] (StrictBind body Stop)
2549 There is just one alternative in the first Select, so we want to
2550 simplify the rhs (I# a) with continuation (StricgtBind body Stop)
2551 Supposing that body is big, we end up with
2552 let $j a = <let x = I# a in body>
2553 in case v of { pn -> case rn of
2554 I# a -> $j a }
2555 This is just what we want because the rn produces a box that
2556 the case rn cancels with.
2557
2558 See Trac #4957 a fuller example.
2559
2560 Note [Case binders and join points]
2561 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2562 Consider this
2563 case (case .. ) of c {
2564 I# c# -> ....c....
2565
2566 If we make a join point with c but not c# we get
2567 $j = \c -> ....c....
2568
2569 But if later inlining scrutines the c, thus
2570
2571 $j = \c -> ... case c of { I# y -> ... } ...
2572
2573 we won't see that 'c' has already been scrutinised. This actually
2574 happens in the 'tabulate' function in wave4main, and makes a significant
2575 difference to allocation.
2576
2577 An alternative plan is this:
2578
2579 $j = \c# -> let c = I# c# in ...c....
2580
2581 but that is bad if 'c' is *not* later scrutinised.
2582
2583 So instead we do both: we pass 'c' and 'c#' , and record in c's inlining
2584 (a stable unfolding) that it's really I# c#, thus
2585
2586 $j = \c# -> \c[=I# c#] -> ...c....
2587
2588 Absence analysis may later discard 'c'.
2589
2590 NB: take great care when doing strictness analysis;
2591 see Note [Lamba-bound unfoldings] in DmdAnal.
2592
2593 Also note that we can still end up passing stuff that isn't used. Before
2594 strictness analysis we have
2595 let $j x y c{=(x,y)} = (h c, ...)
2596 in ...
2597 After strictness analysis we see that h is strict, we end up with
2598 let $j x y c{=(x,y)} = ($wh x y, ...)
2599 and c is unused.
2600
2601 Note [Duplicated env]
2602 ~~~~~~~~~~~~~~~~~~~~~
2603 Some of the alternatives are simplified, but have not been turned into a join point
2604 So they *must* have an zapped subst-env. So we can't use completeNonRecX to
2605 bind the join point, because it might to do PostInlineUnconditionally, and
2606 we'd lose that when zapping the subst-env. We could have a per-alt subst-env,
2607 but zapping it (as we do in mkDupableCont, the Select case) is safe, and
2608 at worst delays the join-point inlining.
2609
2610 Note [Small alternative rhs]
2611 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2612 It is worth checking for a small RHS because otherwise we
2613 get extra let bindings that may cause an extra iteration of the simplifier to
2614 inline back in place. Quite often the rhs is just a variable or constructor.
2615 The Ord instance of Maybe in PrelMaybe.hs, for example, took several extra
2616 iterations because the version with the let bindings looked big, and so wasn't
2617 inlined, but after the join points had been inlined it looked smaller, and so
2618 was inlined.
2619
2620 NB: we have to check the size of rhs', not rhs.
2621 Duplicating a small InAlt might invalidate occurrence information
2622 However, if it *is* dupable, we return the *un* simplified alternative,
2623 because otherwise we'd need to pair it up with an empty subst-env....
2624 but we only have one env shared between all the alts.
2625 (Remember we must zap the subst-env before re-simplifying something).
2626 Rather than do this we simply agree to re-simplify the original (small) thing later.
2627
2628 Note [Funky mkPiTypes]
2629 ~~~~~~~~~~~~~~~~~~~~~~
2630 Notice the funky mkPiTypes. If the contructor has existentials
2631 it's possible that the join point will be abstracted over
2632 type variables as well as term variables.
2633 Example: Suppose we have
2634 data T = forall t. C [t]
2635 Then faced with
2636 case (case e of ...) of
2637 C t xs::[t] -> rhs
2638 We get the join point
2639 let j :: forall t. [t] -> ...
2640 j = /\t \xs::[t] -> rhs
2641 in
2642 case (case e of ...) of
2643 C t xs::[t] -> j t xs
2644
2645 Note [Join point abstraction]
2646 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2647 Join points always have at least one value argument,
2648 for several reasons
2649
2650 * If we try to lift a primitive-typed something out
2651 for let-binding-purposes, we will *caseify* it (!),
2652 with potentially-disastrous strictness results. So
2653 instead we turn it into a function: \v -> e
2654 where v::Void#. The value passed to this function is void,
2655 which generates (almost) no code.
2656
2657 * CPR. We used to say "&& isUnliftedType rhs_ty'" here, but now
2658 we make the join point into a function whenever used_bndrs'
2659 is empty. This makes the join-point more CPR friendly.
2660 Consider: let j = if .. then I# 3 else I# 4
2661 in case .. of { A -> j; B -> j; C -> ... }
2662
2663 Now CPR doesn't w/w j because it's a thunk, so
2664 that means that the enclosing function can't w/w either,
2665 which is a lose. Here's the example that happened in practice:
2666 kgmod :: Int -> Int -> Int
2667 kgmod x y = if x > 0 && y < 0 || x < 0 && y > 0
2668 then 78
2669 else 5
2670
2671 * Let-no-escape. We want a join point to turn into a let-no-escape
2672 so that it is implemented as a jump, and one of the conditions
2673 for LNE is that it's not updatable. In CoreToStg, see
2674 Note [What is a non-escaping let]
2675
2676 * Floating. Since a join point will be entered once, no sharing is
2677 gained by floating out, but something might be lost by doing
2678 so because it might be allocated.
2679
2680 I have seen a case alternative like this:
2681 True -> \v -> ...
2682 It's a bit silly to add the realWorld dummy arg in this case, making
2683 $j = \s v -> ...
2684 True -> $j s
2685 (the \v alone is enough to make CPR happy) but I think it's rare
2686
2687 There's a slight infelicity here: we pass the overall
2688 case_bndr to all the join points if it's used in *any* RHS,
2689 because we don't know its usage in each RHS separately
2690
2691
2692 Note [Duplicating StrictArg]
2693 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2694 The original plan had (where E is a big argument)
2695 e.g. f E [..hole..]
2696 ==> let $j = \a -> f E a
2697 in $j [..hole..]
2698
2699 But this is terrible! Here's an example:
2700 && E (case x of { T -> F; F -> T })
2701 Now, && is strict so we end up simplifying the case with
2702
2703 an ArgOf continuation. If we let-bind it, we get
2704 let $j = \v -> && E v
2705 in simplExpr (case x of { T -> F; F -> T })
2706 (ArgOf (\r -> $j r)
2707 And after simplifying more we get
2708 let $j = \v -> && E v
2709 in case x of { T -> $j F; F -> $j T }
2710 Which is a Very Bad Thing
2711
2712 What we do now is this
2713 f E [..hole..]
2714 ==> let a = E
2715 in f a [..hole..]
2716 Now if the thing in the hole is a case expression (which is when
2717 we'll call mkDupableCont), we'll push the function call into the
2718 branches, which is what we want. Now RULES for f may fire, and
2719 call-pattern specialisation. Here's an example from Trac #3116
2720 go (n+1) (case l of
2721 1 -> bs'
2722 _ -> Chunk p fpc (o+1) (l-1) bs')
2723 If we can push the call for 'go' inside the case, we get
2724 call-pattern specialisation for 'go', which is *crucial* for
2725 this program.
2726
2727 Here is the (&&) example:
2728 && E (case x of { T -> F; F -> T })
2729 ==> let a = E in
2730 case x of { T -> && a F; F -> && a T }
2731 Much better!
2732
2733 Notice that
2734 * Arguments to f *after* the strict one are handled by
2735 the ApplyToVal case of mkDupableCont. Eg
2736 f [..hole..] E
2737
2738 * We can only do the let-binding of E because the function
2739 part of a StrictArg continuation is an explicit syntax
2740 tree. In earlier versions we represented it as a function
2741 (CoreExpr -> CoreEpxr) which we couldn't take apart.
2742
2743 Do *not* duplicate StrictBind and StritArg continuations. We gain
2744 nothing by propagating them into the expressions, and we do lose a
2745 lot.
2746
2747 The desire not to duplicate is the entire reason that
2748 mkDupableCont returns a pair of continuations.
2749
2750 Note [Duplicating StrictBind]
2751 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2752 Unlike StrictArg, there doesn't seem anything to gain from
2753 duplicating a StrictBind continuation, so we don't.
2754
2755
2756 Note [Single-alternative cases]
2757 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2758 This case is just like the ArgOf case. Here's an example:
2759 data T a = MkT !a
2760 ...(MkT (abs x))...
2761 Then we get
2762 case (case x of I# x' ->
2763 case x' <# 0# of
2764 True -> I# (negate# x')
2765 False -> I# x') of y {
2766 DEFAULT -> MkT y
2767 Because the (case x) has only one alternative, we'll transform to
2768 case x of I# x' ->
2769 case (case x' <# 0# of
2770 True -> I# (negate# x')
2771 False -> I# x') of y {
2772 DEFAULT -> MkT y
2773 But now we do *NOT* want to make a join point etc, giving
2774 case x of I# x' ->
2775 let $j = \y -> MkT y
2776 in case x' <# 0# of
2777 True -> $j (I# (negate# x'))
2778 False -> $j (I# x')
2779 In this case the $j will inline again, but suppose there was a big
2780 strict computation enclosing the orginal call to MkT. Then, it won't
2781 "see" the MkT any more, because it's big and won't get duplicated.
2782 And, what is worse, nothing was gained by the case-of-case transform.
2783
2784 So, in circumstances like these, we don't want to build join points
2785 and push the outer case into the branches of the inner one. Instead,
2786 don't duplicate the continuation.
2787
2788 When should we use this strategy? We should not use it on *every*
2789 single-alternative case:
2790 e.g. case (case ....) of (a,b) -> (# a,b #)
2791 Here we must push the outer case into the inner one!
2792 Other choices:
2793
2794 * Match [(DEFAULT,_,_)], but in the common case of Int,
2795 the alternative-filling-in code turned the outer case into
2796 case (...) of y { I# _ -> MkT y }
2797
2798 * Match on single alternative plus (not (isDeadBinder case_bndr))
2799 Rationale: pushing the case inwards won't eliminate the construction.
2800 But there's a risk of
2801 case (...) of y { (a,b) -> let z=(a,b) in ... }
2802 Now y looks dead, but it'll come alive again. Still, this
2803 seems like the best option at the moment.
2804
2805 * Match on single alternative plus (all (isDeadBinder bndrs))
2806 Rationale: this is essentially seq.
2807
2808 * Match when the rhs is *not* duplicable, and hence would lead to a
2809 join point. This catches the disaster-case above. We can test
2810 the *un-simplified* rhs, which is fine. It might get bigger or
2811 smaller after simplification; if it gets smaller, this case might
2812 fire next time round. NB also that we must test contIsDupable
2813 case_cont *too, because case_cont might be big!
2814
2815 HOWEVER: I found that this version doesn't work well, because
2816 we can get let x = case (...) of { small } in ...case x...
2817 When x is inlined into its full context, we find that it was a bad
2818 idea to have pushed the outer case inside the (...) case.
2819
2820 There is a cost to not doing case-of-case; see Trac #10626.
2821
2822 Note [Single-alternative-unlifted]
2823 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2824 Here's another single-alternative where we really want to do case-of-case:
2825
2826 data Mk1 = Mk1 Int# | Mk2 Int#
2827
2828 M1.f =
2829 \r [x_s74 y_s6X]
2830 case
2831 case y_s6X of tpl_s7m {
2832 M1.Mk1 ipv_s70 -> ipv_s70;
2833 M1.Mk2 ipv_s72 -> ipv_s72;
2834 }
2835 of
2836 wild_s7c
2837 { __DEFAULT ->
2838 case
2839 case x_s74 of tpl_s7n {
2840 M1.Mk1 ipv_s77 -> ipv_s77;
2841 M1.Mk2 ipv_s79 -> ipv_s79;
2842 }
2843 of
2844 wild1_s7b
2845 { __DEFAULT -> ==# [wild1_s7b wild_s7c];
2846 };
2847 };
2848
2849 So the outer case is doing *nothing at all*, other than serving as a
2850 join-point. In this case we really want to do case-of-case and decide
2851 whether to use a real join point or just duplicate the continuation:
2852
2853 let $j s7c = case x of
2854 Mk1 ipv77 -> (==) s7c ipv77
2855 Mk1 ipv79 -> (==) s7c ipv79
2856 in
2857 case y of
2858 Mk1 ipv70 -> $j ipv70
2859 Mk2 ipv72 -> $j ipv72
2860
2861 Hence: check whether the case binder's type is unlifted, because then
2862 the outer case is *not* a seq.
2863
2864 ************************************************************************
2865 * *
2866 Unfoldings
2867 * *
2868 ************************************************************************
2869 -}
2870
2871 simplLetUnfolding :: SimplEnv-> TopLevelFlag
2872 -> InId
2873 -> OutExpr
2874 -> Unfolding -> SimplM Unfolding
2875 simplLetUnfolding env top_lvl id new_rhs unf
2876 | isStableUnfolding unf
2877 = simplUnfolding env top_lvl id unf
2878 | otherwise
2879 = bottoming `seq` -- See Note [Force bottoming field]
2880 do { dflags <- getDynFlags
2881 ; return (mkUnfolding dflags InlineRhs (isTopLevel top_lvl) bottoming new_rhs) }
2882 -- We make an unfolding *even for loop-breakers*.
2883 -- Reason: (a) It might be useful to know that they are WHNF
2884 -- (b) In TidyPgm we currently assume that, if we want to
2885 -- expose the unfolding then indeed we *have* an unfolding
2886 -- to expose. (We could instead use the RHS, but currently
2887 -- we don't.) The simple thing is always to have one.
2888 where
2889 bottoming = isBottomingId id
2890
2891 simplUnfolding :: SimplEnv-> TopLevelFlag -> InId -> Unfolding -> SimplM Unfolding
2892 -- Note [Setting the new unfolding]
2893 simplUnfolding env top_lvl id unf
2894 = case unf of
2895 NoUnfolding -> return unf
2896 OtherCon {} -> return unf
2897
2898 DFunUnfolding { df_bndrs = bndrs, df_con = con, df_args = args }
2899 -> do { (env', bndrs') <- simplBinders rule_env bndrs
2900 ; args' <- mapM (simplExpr env') args
2901 ; return (mkDFunUnfolding bndrs' con args') }
2902
2903 CoreUnfolding { uf_tmpl = expr, uf_src = src, uf_guidance = guide }
2904 | isStableSource src
2905 -> do { expr' <- simplExpr rule_env expr
2906 ; case guide of
2907 UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok } -- Happens for INLINE things
2908 -> let guide' = UnfWhen { ug_arity = arity, ug_unsat_ok = sat_ok
2909 , ug_boring_ok = inlineBoringOk expr' }
2910 -- Refresh the boring-ok flag, in case expr'
2911 -- has got small. This happens, notably in the inlinings
2912 -- for dfuns for single-method classes; see
2913 -- Note [Single-method classes] in TcInstDcls.
2914 -- A test case is Trac #4138
2915 in return (mkCoreUnfolding src is_top_lvl expr' guide')
2916 -- See Note [Top-level flag on inline rules] in CoreUnfold
2917
2918 _other -- Happens for INLINABLE things
2919 -> bottoming `seq` -- See Note [Force bottoming field]
2920 do { dflags <- getDynFlags
2921 ; return (mkUnfolding dflags src is_top_lvl bottoming expr') } }
2922 -- If the guidance is UnfIfGoodArgs, this is an INLINABLE
2923 -- unfolding, and we need to make sure the guidance is kept up
2924 -- to date with respect to any changes in the unfolding.
2925
2926 | otherwise -> return noUnfolding -- Discard unstable unfoldings
2927 where
2928 bottoming = isBottomingId id
2929 is_top_lvl = isTopLevel top_lvl
2930 act = idInlineActivation id
2931 rule_env = updMode (updModeForStableUnfoldings act) env
2932 -- See Note [Simplifying inside stable unfoldings] in SimplUtils
2933
2934 {-
2935 Note [Force bottoming field]
2936 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2937 We need to force bottoming, or the new unfolding holds
2938 on to the old unfolding (which is part of the id).
2939
2940 Note [Setting the new unfolding]
2941 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2942 * If there's an INLINE pragma, we simplify the RHS gently. Maybe we
2943 should do nothing at all, but simplifying gently might get rid of
2944 more crap.
2945
2946 * If not, we make an unfolding from the new RHS. But *only* for
2947 non-loop-breakers. Making loop breakers not have an unfolding at all
2948 means that we can avoid tests in exprIsConApp, for example. This is
2949 important: if exprIsConApp says 'yes' for a recursive thing, then we
2950 can get into an infinite loop
2951
2952 If there's an stable unfolding on a loop breaker (which happens for
2953 INLINEABLE), we hang on to the inlining. It's pretty dodgy, but the
2954 user did say 'INLINE'. May need to revisit this choice.
2955
2956 ************************************************************************
2957 * *
2958 Rules
2959 * *
2960 ************************************************************************
2961
2962 Note [Rules in a letrec]
2963 ~~~~~~~~~~~~~~~~~~~~~~~~
2964 After creating fresh binders for the binders of a letrec, we
2965 substitute the RULES and add them back onto the binders; this is done
2966 *before* processing any of the RHSs. This is important. Manuel found
2967 cases where he really, really wanted a RULE for a recursive function
2968 to apply in that function's own right-hand side.
2969
2970 See Note [Loop breaking and RULES] in OccAnal.
2971 -}
2972
2973 addBndrRules :: SimplEnv -> InBndr -> OutBndr -> SimplM (SimplEnv, OutBndr)
2974 -- Rules are added back into the bin
2975 addBndrRules env in_id out_id
2976 | null old_rules
2977 = return (env, out_id)
2978 | otherwise
2979 = do { new_rules <- simplRules env (Just (idName out_id)) old_rules
2980 ; let final_id = out_id `setIdSpecialisation` mkRuleInfo new_rules
2981 ; return (modifyInScope env final_id, final_id) }
2982 where
2983 old_rules = ruleInfoRules (idSpecialisation in_id)
2984
2985 simplRules :: SimplEnv -> Maybe Name -> [CoreRule] -> SimplM [CoreRule]
2986 simplRules env mb_new_nm rules
2987 = mapM simpl_rule rules
2988 where
2989 simpl_rule rule@(BuiltinRule {})
2990 = return rule
2991
2992 simpl_rule rule@(Rule { ru_bndrs = bndrs, ru_args = args
2993 , ru_fn = fn_name, ru_rhs = rhs })
2994 = do { (env', bndrs') <- simplBinders env bndrs
2995 ; let rule_env = updMode updModeForRules env'
2996 ; args' <- mapM (simplExpr rule_env) args
2997 ; rhs' <- simplExpr rule_env rhs
2998 ; return (rule { ru_bndrs = bndrs'
2999 , ru_fn = mb_new_nm `orElse` fn_name
3000 , ru_args = args'
3001 , ru_rhs = rhs' }) }